{
  "status": "reviewed_complete",
  "generated_unix": 1790046344.017151,
  "metrics": {
    "hf-A0": {
      "engine": "hf",
      "candidate": "A0",
      "records": 160,
      "failed_requests": [],
      "final_document_expected": 128,
      "final_document_present": 128,
      "final_document_failed": 0,
      "quality_complete": true,
      "nll": 1.93365006070158,
      "ppl": 6.914703322333733,
      "scored_tokens": 130944,
      "domains": {
        "ko": {
          "documents": 64,
          "nll": 2.1870921051688677,
          "ppl": 8.909268195497868
        },
        "en": {
          "documents": 32,
          "nll": 2.2387158178222726,
          "ppl": 9.38127628088293
        },
        "code": {
          "documents": 32,
          "nll": 1.1217002146463124,
          "ppl": 3.0700695456426317
        }
      },
      "delta_nll": {
        "delta": 0.0820362158593078,
        "ci95": [
          0.07676074866274643,
          0.08754806751121319
        ],
        "resampling_unit": "paired document",
        "replicates": 2000
      },
      "ppl_ratio_vs_bf16": {
        "ratio": 1.0854951212562869,
        "ci95": [
          1.0797837057905217,
          1.0914947286171375
        ]
      },
      "document_delta_tail": {
        "mean": 0.08203621585930782,
        "median": 0.0820228557083264,
        "p95": 0.13523329350722266,
        "p99": 0.1744629132689386,
        "max": 0.1855487040410643
      },
      "worst20_documents": [
        {
          "id": "final-109",
          "domain": "en",
          "delta_nll": 0.1855487040410643,
          "nll": 2.280070516487487
        },
        {
          "id": "final-049",
          "domain": "code",
          "delta_nll": 0.17658769676529185,
          "nll": 0.6997955281247374
        },
        {
          "id": "final-035",
          "domain": "ko",
          "delta_nll": 0.16871812826027968,
          "nll": 2.2584909539768074
        },
        {
          "id": "final-103",
          "domain": "en",
          "delta_nll": 0.15268722470787846,
          "nll": 2.2268263955736205
        },
        {
          "id": "final-009",
          "domain": "ko",
          "delta_nll": 0.14515656390148174,
          "nll": 2.3813408164567846
        },
        {
          "id": "final-075",
          "domain": "ko",
          "delta_nll": 0.1392123277352819,
          "nll": 2.0740082350178084
        },
        {
          "id": "final-018",
          "domain": "ko",
          "delta_nll": 0.13586595634095014,
          "nll": 2.4304069451922197
        },
        {
          "id": "final-003",
          "domain": "ko",
          "delta_nll": 0.13405834824458596,
          "nll": 1.5526132835321063
        },
        {
          "id": "final-063",
          "domain": "ko",
          "delta_nll": 0.13208126578978896,
          "nll": 2.06155704170145
        },
        {
          "id": "final-012",
          "domain": "code",
          "delta_nll": 0.12857145065203207,
          "nll": 0.4027091927775423
        },
        {
          "id": "final-001",
          "domain": "code",
          "delta_nll": 0.12701752365276378,
          "nll": 0.32724712274873946
        },
        {
          "id": "final-036",
          "domain": "ko",
          "delta_nll": 0.1217827195645777,
          "nll": 2.393973104415401
        },
        {
          "id": "final-039",
          "domain": "ko",
          "delta_nll": 0.12090000122756436,
          "nll": 2.3648213846126493
        },
        {
          "id": "final-032",
          "domain": "ko",
          "delta_nll": 0.11797917833076545,
          "nll": 1.9544753124054344
        },
        {
          "id": "final-033",
          "domain": "ko",
          "delta_nll": 0.11529606644825269,
          "nll": 2.4616481048270753
        },
        {
          "id": "final-113",
          "domain": "ko",
          "delta_nll": 0.11503100092226926,
          "nll": 2.366507243084651
        },
        {
          "id": "final-046",
          "domain": "code",
          "delta_nll": 0.11190791237272708,
          "nll": 1.3897814177348118
        },
        {
          "id": "final-047",
          "domain": "ko",
          "delta_nll": 0.11058930893913858,
          "nll": 1.6414782280796085
        },
        {
          "id": "final-019",
          "domain": "en",
          "delta_nll": 0.11052950805932546,
          "nll": 2.0117747268602177
        },
        {
          "id": "final-107",
          "domain": "ko",
          "delta_nll": 0.10837323178527175,
          "nll": 2.1305793825598873
        }
      ],
      "token_delta_tail": {
        "mean": 0.0820362158593078,
        "median": 0.004975318908691406,
        "p95": 0.8666933059692374,
        "p99": 1.8084007644653302,
        "max": 15.288114547729492
      },
      "sampled_full_vocab_kl": {
        "mean": 0.0970104158181517,
        "median": 0.042478140968257146,
        "p95": 0.29837098349910446,
        "p99": 0.7911814699498793,
        "max": 17.646372187274768
      },
      "kl_positions": 4096,
      "sampled_top1_agreement": 0.87255859375,
      "tasks": {},
      "speed": {},
      "speed_source": "embedded workload-tail run; dedicated benchmark pending",
      "block_diagnostics": {
        "scope": "six fixed validation documents, 256 tokens, cumulative error",
        "layer_mean_relative_l2": {
          "0": 0.12903799457971124,
          "1": 0.1571070468040269,
          "2": 0.22566349600917013,
          "3": 0.21133631629360275,
          "4": 0.2285544830943772,
          "5": 0.23136062024308487,
          "6": 0.23393891724912053,
          "7": 0.2467944833779129,
          "8": 0.24475813253585876,
          "9": 0.24711447521186036,
          "10": 0.24167888034169704,
          "11": 0.25611256562120915,
          "12": 0.2604246348639825,
          "13": 0.2591232188818453,
          "14": 0.2608755204209323,
          "15": 0.27006129648057214,
          "16": 0.2737835062148852,
          "17": 0.2643662372116999,
          "18": 0.2542052663037596,
          "19": 0.2652235503473812,
          "20": 0.2567186581435374,
          "21": 0.2535158345391283,
          "22": 0.25320365483126345,
          "23": 0.2543835751826935,
          "24": 0.2530360601866923,
          "25": 0.25580827084285024,
          "26": 0.26490951443777777,
          "27": 0.27515065090187246,
          "28": 0.2654177558413169,
          "29": 0.271302181349774,
          "30": 0.2837576174019451,
          "31": 0.2844029517845293
        }
      }
    },
    "hf-B0": {
      "engine": "hf",
      "candidate": "B0",
      "records": 160,
      "failed_requests": [],
      "final_document_expected": 128,
      "final_document_present": 128,
      "final_document_failed": 0,
      "quality_complete": true,
      "nll": 1.851613844842272,
      "ppl": 6.370091571052912,
      "scored_tokens": 130944,
      "domains": {
        "ko": {
          "documents": 64,
          "nll": 2.0939926444639325,
          "ppl": 8.117259883591567
        },
        "en": {
          "documents": 32,
          "nll": 2.1686982321366304,
          "ppl": 8.746890205711475
        },
        "code": {
          "documents": 32,
          "nll": 1.0497718583045934,
          "ppl": 2.8569992430549704
        }
      },
      "delta_nll": {
        "delta": 0.0,
        "ci95": [
          0.0,
          0.0
        ],
        "resampling_unit": "paired document",
        "replicates": 2000
      },
      "ppl_ratio_vs_bf16": {
        "ratio": 1.0,
        "ci95": [
          1.0,
          1.0
        ]
      },
      "document_delta_tail": {
        "mean": 0.0,
        "median": 0.0,
        "p95": 0.0,
        "p99": 0.0,
        "max": 0.0
      },
      "worst20_documents": [
        {
          "id": "final-000",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 2.210419322039025
        },
        {
          "id": "final-001",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 0.20022959909597565
        },
        {
          "id": "final-002",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.8950367085744908
        },
        {
          "id": "final-003",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.4185549352875204
        },
        {
          "id": "final-004",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 1.1785066153413506
        },
        {
          "id": "final-005",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 2.2057225897631456
        },
        {
          "id": "final-006",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 0.7938262038916087
        },
        {
          "id": "final-007",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.9395612649554033
        },
        {
          "id": "final-008",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 2.2966494592869853
        },
        {
          "id": "final-009",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.2361842525553026
        },
        {
          "id": "final-010",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.424708617164709
        },
        {
          "id": "final-011",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.568956426394883
        },
        {
          "id": "final-012",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 0.27413774212551023
        },
        {
          "id": "final-013",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 2.472125922246879
        },
        {
          "id": "final-014",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.6370440163337356
        },
        {
          "id": "final-015",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.215205994286262
        },
        {
          "id": "final-016",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.339994630039845
        },
        {
          "id": "final-017",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 1.0864892625855327
        },
        {
          "id": "final-018",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.2945409888512693
        },
        {
          "id": "final-019",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 1.9012452188008924
        }
      ],
      "token_delta_tail": {
        "mean": 0.0,
        "median": 0.0,
        "p95": 0.0,
        "p99": 0.0,
        "max": 0.0
      },
      "sampled_full_vocab_kl": null,
      "kl_positions": 0,
      "tasks": {},
      "speed": {},
      "speed_source": "embedded workload-tail run; dedicated benchmark pending",
      "block_diagnostics": {
        "scope": "six fixed validation documents, 256 tokens, cumulative error",
        "layer_mean_relative_l2": {
          "0": 0.0,
          "1": 0.0,
          "2": 0.0,
          "3": 0.0,
          "4": 0.0,
          "5": 0.0,
          "6": 0.0,
          "7": 0.0,
          "8": 0.0,
          "9": 0.0,
          "10": 0.0,
          "11": 0.0,
          "12": 0.0,
          "13": 0.0,
          "14": 0.0,
          "15": 0.0,
          "16": 0.0,
          "17": 0.0,
          "18": 0.0,
          "19": 0.0,
          "20": 0.0,
          "21": 0.0,
          "22": 0.0,
          "23": 0.0,
          "24": 0.0,
          "25": 0.0,
          "26": 0.0,
          "27": 0.0,
          "28": 0.0,
          "29": 0.0,
          "30": 0.0,
          "31": 0.0
        }
      }
    },
    "hf-G0": {
      "engine": "hf",
      "candidate": "G0",
      "records": 160,
      "failed_requests": [],
      "final_document_expected": 128,
      "final_document_present": 128,
      "final_document_failed": 0,
      "quality_complete": true,
      "nll": 1.8885224431476402,
      "ppl": 6.609595409550301,
      "scored_tokens": 130944,
      "domains": {
        "ko": {
          "documents": 64,
          "nll": 2.132348335019421,
          "ppl": 8.434650958668149
        },
        "en": {
          "documents": 32,
          "nll": 2.2070701874479406,
          "ppl": 9.08904813830095
        },
        "code": {
          "documents": 32,
          "nll": 1.082322915103778,
          "ppl": 2.9515277424747435
        }
      },
      "delta_nll": {
        "delta": 0.036908598305368,
        "ci95": [
          0.03381318624021225,
          0.040242480689282815
        ],
        "resampling_unit": "paired document",
        "replicates": 2000
      },
      "ppl_ratio_vs_bf16": {
        "ratio": 1.0375981782719963,
        "ci95": [
          1.0343913501399735,
          1.0410631813070161
        ]
      },
      "document_delta_tail": {
        "mean": 0.036908598305368,
        "median": 0.03632262118983712,
        "p95": 0.06311846073072325,
        "p99": 0.10133215486595487,
        "max": 0.13963556429507795
      },
      "worst20_documents": [
        {
          "id": "final-049",
          "domain": "code",
          "delta_nll": 0.13963556429507795,
          "nll": 0.6628433956545235
        },
        {
          "id": "final-103",
          "domain": "en",
          "delta_nll": 0.10965255185422776,
          "nll": 2.18379172271997
        },
        {
          "id": "final-036",
          "domain": "ko",
          "delta_nll": 0.07883626671247584,
          "nll": 2.3510266515632994
        },
        {
          "id": "final-035",
          "domain": "ko",
          "delta_nll": 0.07751580999976612,
          "nll": 2.167288635716294
        },
        {
          "id": "final-013",
          "domain": "en",
          "delta_nll": 0.07282334175743782,
          "nll": 2.5449492640043165
        },
        {
          "id": "final-046",
          "domain": "code",
          "delta_nll": 0.0649968172448122,
          "nll": 1.3428703226068968
        },
        {
          "id": "final-063",
          "domain": "ko",
          "delta_nll": 0.06427158568262005,
          "nll": 1.9937473615942813
        },
        {
          "id": "final-019",
          "domain": "en",
          "delta_nll": 0.06097694296291497,
          "nll": 1.9622221617638074
        },
        {
          "id": "final-047",
          "domain": "ko",
          "delta_nll": 0.058374237105294054,
          "nll": 1.589263156245764
        },
        {
          "id": "final-073",
          "domain": "en",
          "delta_nll": 0.05784474649736958,
          "nll": 2.2561646747682333
        },
        {
          "id": "final-083",
          "domain": "code",
          "delta_nll": 0.05595014708249567,
          "nll": 1.1111934219986812
        },
        {
          "id": "final-024",
          "domain": "ko",
          "delta_nll": 0.05479627014488302,
          "nll": 2.1866303319921707
        },
        {
          "id": "final-124",
          "domain": "ko",
          "delta_nll": 0.05446133282643725,
          "nll": 2.1020163781249046
        },
        {
          "id": "final-030",
          "domain": "ko",
          "delta_nll": 0.05397154415574358,
          "nll": 1.5014877142211676
        },
        {
          "id": "final-039",
          "domain": "ko",
          "delta_nll": 0.05318891314583906,
          "nll": 2.297110296530924
        },
        {
          "id": "final-032",
          "domain": "ko",
          "delta_nll": 0.053079588322345814,
          "nll": 1.8895757223970147
        },
        {
          "id": "final-003",
          "domain": "ko",
          "delta_nll": 0.052185427758001515,
          "nll": 1.4707403630455218
        },
        {
          "id": "final-091",
          "domain": "en",
          "delta_nll": 0.05165458168335558,
          "nll": 2.3481659865915367
        },
        {
          "id": "final-101",
          "domain": "en",
          "delta_nll": 0.05118013756715657,
          "nll": 2.287601511033283
        },
        {
          "id": "final-126",
          "domain": "ko",
          "delta_nll": 0.05110409252804395,
          "nll": 2.4185564548272427
        }
      ],
      "token_delta_tail": {
        "mean": 0.036908598305368,
        "median": 0.001071929931640625,
        "p95": 0.5307848930358877,
        "p99": 1.1659850120544428,
        "max": 9.068389892578125
      },
      "sampled_full_vocab_kl": {
        "mean": 0.044962490019749626,
        "median": 0.016390751744274946,
        "p95": 0.14909470627543128,
        "p99": 0.5072163384931685,
        "max": 4.771928180072513
      },
      "kl_positions": 4096,
      "sampled_top1_agreement": 0.916748046875,
      "tasks": {},
      "speed": {},
      "speed_source": "embedded workload-tail run; dedicated benchmark pending",
      "block_diagnostics": {
        "scope": "six fixed validation documents, 256 tokens, cumulative error",
        "layer_mean_relative_l2": {
          "0": 0.061502795778974495,
          "1": 0.08756348352111992,
          "2": 0.13211320323248357,
          "3": 0.13194254501285904,
          "4": 0.1496387714938382,
          "5": 0.15543279143382632,
          "6": 0.15926372939334799,
          "7": 0.1705592222909812,
          "8": 0.17304152740954593,
          "9": 0.1773070721051936,
          "10": 0.17448204471676795,
          "11": 0.1832057436036161,
          "12": 0.18618751560964827,
          "13": 0.18606091406986847,
          "14": 0.1868976624418859,
          "15": 0.19186519409745897,
          "16": 0.19513437758205152,
          "17": 0.18753628823447765,
          "18": 0.17669351825942817,
          "19": 0.1808796693581447,
          "20": 0.17552757239029024,
          "21": 0.17439488314077153,
          "22": 0.17805623219860633,
          "23": 0.17567599658009236,
          "24": 0.17677755667138217,
          "25": 0.1811603875257962,
          "26": 0.19488573127430076,
          "27": 0.19370679826542161,
          "28": 0.18927246627180824,
          "29": 0.1957854662306057,
          "30": 0.20543273546404092,
          "31": 0.20172448927764497
        }
      }
    },
    "hf-R0": {
      "engine": "hf",
      "candidate": "R0",
      "records": 160,
      "failed_requests": [],
      "final_document_expected": 128,
      "final_document_present": 128,
      "final_document_failed": 0,
      "quality_complete": true,
      "nll": 1.9407359235214465,
      "ppl": 6.963873963990896,
      "scored_tokens": 130944,
      "domains": {
        "ko": {
          "documents": 64,
          "nll": 2.198646108003539,
          "ppl": 9.012802873688717
        },
        "en": {
          "documents": 32,
          "nll": 2.2330728216371694,
          "ppl": 9.328486859951811
        },
        "code": {
          "documents": 32,
          "nll": 1.1325786564415385,
          "ppl": 3.103649435788002
        }
      },
      "delta_nll": {
        "delta": 0.08912207867917427,
        "ci95": [
          0.08320109213785225,
          0.09499921878273593
        ],
        "resampling_unit": "paired document",
        "replicates": 2000
      },
      "ppl_ratio_vs_bf16": {
        "ratio": 1.0932141063146188,
        "ci95": [
          1.0867603255336724,
          1.0996579960539563
        ]
      },
      "document_delta_tail": {
        "mean": 0.08912207867917427,
        "median": 0.08879938806140411,
        "p95": 0.14887570314043774,
        "p99": 0.16015912470812904,
        "max": 0.19972954560817743
      },
      "worst20_documents": [
        {
          "id": "final-049",
          "domain": "code",
          "delta_nll": 0.19972954560817743,
          "nll": 0.7229373769676231
        },
        {
          "id": "final-063",
          "domain": "ko",
          "delta_nll": 0.16074700602571518,
          "nll": 2.0902227819373764
        },
        {
          "id": "final-103",
          "domain": "en",
          "delta_nll": 0.15856966781243317,
          "nll": 2.2327088386781755
        },
        {
          "id": "final-003",
          "domain": "ko",
          "delta_nll": 0.15642737689722086,
          "nll": 1.5749823121847413
        },
        {
          "id": "final-001",
          "domain": "code",
          "delta_nll": 0.15544412842244346,
          "nll": 0.35567372751841914
        },
        {
          "id": "final-035",
          "domain": "ko",
          "delta_nll": 0.15013684415397757,
          "nll": 2.239909669870505
        },
        {
          "id": "final-085",
          "domain": "ko",
          "delta_nll": 0.14948467728213014,
          "nll": 2.78201802385168
        },
        {
          "id": "final-012",
          "domain": "code",
          "delta_nll": 0.14774475116300909,
          "nll": 0.42188249328851934
        },
        {
          "id": "final-018",
          "domain": "ko",
          "delta_nll": 0.14653743863222304,
          "nll": 2.4410784274834927
        },
        {
          "id": "final-033",
          "domain": "ko",
          "delta_nll": 0.13925958728510612,
          "nll": 2.4856116256639287
        },
        {
          "id": "final-047",
          "domain": "ko",
          "delta_nll": 0.13845634553672514,
          "nll": 1.6693452646771951
        },
        {
          "id": "final-036",
          "domain": "ko",
          "delta_nll": 0.13749841837356283,
          "nll": 2.4096888032243866
        },
        {
          "id": "final-075",
          "domain": "ko",
          "delta_nll": 0.13698122205389326,
          "nll": 2.0717771293364198
        },
        {
          "id": "final-126",
          "domain": "ko",
          "delta_nll": 0.12869443734953015,
          "nll": 2.4961467996487285
        },
        {
          "id": "final-118",
          "domain": "ko",
          "delta_nll": 0.12809968017995765,
          "nll": 2.2436153443561153
        },
        {
          "id": "final-052",
          "domain": "code",
          "delta_nll": 0.12747464221943158,
          "nll": 1.2281017331433786
        },
        {
          "id": "final-009",
          "domain": "ko",
          "delta_nll": 0.12590000380052732,
          "nll": 2.36208425635583
        },
        {
          "id": "final-057",
          "domain": "ko",
          "delta_nll": 0.12252216395045305,
          "nll": 2.6040335633771043
        },
        {
          "id": "final-039",
          "domain": "ko",
          "delta_nll": 0.12172212349005108,
          "nll": 2.365643506875136
        },
        {
          "id": "final-046",
          "domain": "code",
          "delta_nll": 0.12160308852806352,
          "nll": 1.3994765938901481
        }
      ],
      "token_delta_tail": {
        "mean": 0.08912207867917427,
        "median": 0.00594329833984375,
        "p95": 0.9231513023376459,
        "p99": 1.928824300765982,
        "max": 11.330570220947266
      },
      "sampled_full_vocab_kl": {
        "mean": 0.1059899188939109,
        "median": 0.046785345441875806,
        "p95": 0.33871690559947343,
        "p99": 0.9872271586607808,
        "max": 11.07881304550959
      },
      "kl_positions": 4096,
      "sampled_top1_agreement": 0.865234375,
      "tasks": {},
      "speed": {},
      "speed_source": "embedded workload-tail run; dedicated benchmark pending",
      "block_diagnostics": {
        "scope": "six fixed validation documents, 256 tokens, cumulative error",
        "layer_mean_relative_l2": {
          "0": 0.1295213972207483,
          "1": 0.1625967985442728,
          "2": 0.23553434934653958,
          "3": 0.21979621509447492,
          "4": 0.2376477745018263,
          "5": 0.24059236440198561,
          "6": 0.24654153530486314,
          "7": 0.260415532026211,
          "8": 0.2598454006934941,
          "9": 0.2624094256982629,
          "10": 0.25614311139116425,
          "11": 0.26872622260569584,
          "12": 0.2726684851022761,
          "13": 0.27115011713649545,
          "14": 0.27342810173514115,
          "15": 0.2818072557633838,
          "16": 0.2858158542715014,
          "17": 0.2768645915943102,
          "18": 0.2662182572484813,
          "19": 0.2752910905346593,
          "20": 0.266664676018941,
          "21": 0.2632310743354779,
          "22": 0.2639017946895043,
          "23": 0.26235592497209725,
          "24": 0.2594317109714475,
          "25": 0.2626580358224639,
          "26": 0.2774713093128148,
          "27": 0.27985080119603567,
          "28": 0.27082239671685315,
          "29": 0.2777480940112989,
          "30": 0.28839128452507495,
          "31": 0.2912248048843746
        }
      }
    },
    "native-B0": {
      "engine": "native",
      "candidate": "B0",
      "records": 4714,
      "failed_requests": [],
      "final_document_expected": 128,
      "final_document_present": 128,
      "final_document_failed": 0,
      "quality_complete": true,
      "nll": 1.8514976427891694,
      "ppl": 6.369351396339647,
      "scored_tokens": 130944,
      "domains": {
        "ko": {
          "documents": 64,
          "nll": 2.0938041075234968,
          "ppl": 8.11572962450811
        },
        "en": {
          "documents": 32,
          "nll": 2.168783994885714,
          "ppl": 8.747640395230148
        },
        "code": {
          "documents": 32,
          "nll": 1.0495983612239705,
          "ppl": 2.856503605024078
        }
      },
      "delta_nll": {
        "delta": 0.0,
        "ci95": [
          0.0,
          0.0
        ],
        "resampling_unit": "paired document",
        "replicates": 2000
      },
      "ppl_ratio_vs_bf16": {
        "ratio": 1.0,
        "ci95": [
          1.0,
          1.0
        ]
      },
      "document_delta_tail": {
        "mean": 0.0,
        "median": 0.0,
        "p95": 0.0,
        "p99": 0.0,
        "max": 0.0
      },
      "worst20_documents": [
        {
          "id": "final-000",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 2.210358508974271
        },
        {
          "id": "final-001",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 0.19993765938142033
        },
        {
          "id": "final-002",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.8954136529580778
        },
        {
          "id": "final-003",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.4191517381421672
        },
        {
          "id": "final-004",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 1.1780465211209783
        },
        {
          "id": "final-005",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 2.2044185917991137
        },
        {
          "id": "final-006",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 0.7937105457941359
        },
        {
          "id": "final-007",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.94125673385177
        },
        {
          "id": "final-008",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 2.2967991607329763
        },
        {
          "id": "final-009",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.236059863709279
        },
        {
          "id": "final-010",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.423272220184445
        },
        {
          "id": "final-011",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.56812581664109
        },
        {
          "id": "final-012",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 0.27485824072762477
        },
        {
          "id": "final-013",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 2.4718865563630583
        },
        {
          "id": "final-014",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.6355058926572874
        },
        {
          "id": "final-015",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.2157928098594526
        },
        {
          "id": "final-016",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.339425756856219
        },
        {
          "id": "final-017",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 1.0859001591590602
        },
        {
          "id": "final-018",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.2943346452131936
        },
        {
          "id": "final-019",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 1.9004068108754806
        }
      ],
      "token_delta_tail": {
        "mean": 0.0,
        "median": 0.0,
        "p95": 0.0,
        "p99": 0.0,
        "max": 0.0
      },
      "sampled_full_vocab_kl": null,
      "kl_positions": 0,
      "tasks": {
        "kmmlu": {
          "n": 500,
          "accuracy": 0.482,
          "failures": 0,
          "subject_scores": {
            "Accounting": {
              "n": 12,
              "accuracy": 0.5833333333333334
            },
            "Agricultural-Sciences": {
              "n": 12,
              "accuracy": 0.5
            },
            "Aviation-Engineering-and-Maintenance": {
              "n": 12,
              "accuracy": 0.5
            },
            "Biology": {
              "n": 12,
              "accuracy": 0.3333333333333333
            },
            "Chemical-Engineering": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Chemistry": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Civil-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Computer-Science": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Construction": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Criminal-Law": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Ecology": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Economics": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Education": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Electrical-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Electronics-Engineering": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Energy-Management": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Environmental-Science": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Fashion": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Food-Processing": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Gas-Technology-and-Engineering": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Geomatics": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Health": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Industrial-Engineer": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Information-Technology": {
              "n": 11,
              "accuracy": 1.0
            },
            "Interior-Architecture-and-Design": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Korean-History": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Law": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Machine-Design-and-Manufacturing": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Management": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Maritime-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Marketing": {
              "n": 11,
              "accuracy": 0.9090909090909091
            },
            "Materials-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Math": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Mechanical-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Nondestructive-Testing": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Patent": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Political-Science-and-Sociology": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Psychology": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Public-Safety": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Railway-and-Automotive-Engineering": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Real-Estate": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Refrigerating-Machinery": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Social-Welfare": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Taxation": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Telecommunications-and-Wireless-Technology": {
              "n": 11,
              "accuracy": 0.6363636363636364
            }
          },
          "macro_accuracy": 0.48215488215488206,
          "delta_accuracy": {
            "delta": 0.0,
            "ci95": [
              0.0,
              0.0
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [],
          "improvements": []
        },
        "arc": {
          "n": 500,
          "accuracy": 0.518,
          "failures": 0,
          "accuracy_norm": 0.562,
          "delta_accuracy": {
            "delta": 0.0,
            "ci95": [
              0.0,
              0.0
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [],
          "improvements": [],
          "delta_accuracy_norm": {
            "delta": 0.0,
            "ci95": [
              0.0,
              0.0
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          }
        },
        "ifeval": {
          "n": 541,
          "accuracy": 0.8262476894639557,
          "failures": 0,
          "instruction_strict": 0.8776978417266187,
          "truncated": 13,
          "delta_accuracy": {
            "delta": 0.0,
            "ci95": [
              0.0,
              0.0
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [],
          "improvements": []
        }
      },
      "speed": {
        "512": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.0382353,
            "min": 0.0380698,
            "max": 0.0384302
          },
          "prefill_seconds": {
            "median": 0.0382093,
            "min": 0.0380427,
            "max": 0.0384085
          },
          "decode_seconds": {
            "median": 1.6359083,
            "min": 1.5475906,
            "max": 1.6526967
          },
          "seconds": {
            "median": 1.6739789,
            "min": 1.5858268,
            "max": 1.6911273
          },
          "prefill_tokens_per_second": {
            "median": 13399.87908702855,
            "min": 13330.382597602094,
            "max": 13458.56103799152
          },
          "decode_tokens_per_second": {
            "median": 155.87670775923075,
            "min": 154.29328321403437,
            "max": 164.77225953685684
          },
          "peak_device_vram_bytes": 12694835200
        },
        "4096": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.3036894,
            "min": 0.3015319,
            "max": 0.3110883
          },
          "prefill_seconds": {
            "median": 0.3036635,
            "min": 0.3015077,
            "max": 0.311065
          },
          "decode_seconds": {
            "median": 1.6379478,
            "min": 1.569075,
            "max": 1.6604678
          },
          "seconds": {
            "median": 1.9427159,
            "min": 1.8709703,
            "max": 1.9620004
          },
          "prefill_tokens_per_second": {
            "median": 13488.614864809238,
            "min": 13167.665921913427,
            "max": 13585.059353376382
          },
          "decode_tokens_per_second": {
            "median": 155.68261699182355,
            "min": 153.57118036254602,
            "max": 162.51613211605564
          },
          "peak_device_vram_bytes": 12694835200
        }
      },
      "speed_source": "dedicated serial benchmark",
      "runtime_memory_breakdown": {
        "source": "llama.cpp CUDA0 buffer log",
        "model_mib": 8023.67,
        "kv_mib": 256.0,
        "recurrent_state_mib": 50.25,
        "compute_mib": 495.0
      },
      "block_diagnostics": {
        "scope": "six fixed validation documents, 256 tokens, cumulative error",
        "layer_mean_relative_l2": {
          "0": 0.0,
          "1": 0.0,
          "2": 0.0,
          "3": 0.0,
          "4": 0.0,
          "5": 0.0,
          "6": 0.0,
          "7": 0.0,
          "8": 0.0,
          "9": 0.0,
          "10": 0.0,
          "11": 0.0,
          "12": 0.0,
          "13": 0.0,
          "14": 0.0,
          "15": 0.0,
          "16": 0.0,
          "17": 0.0,
          "18": 0.0,
          "19": 0.0,
          "20": 0.0,
          "21": 0.0,
          "22": 0.0,
          "23": 0.0,
          "24": 0.0,
          "25": 0.0,
          "26": 0.0,
          "27": 0.0,
          "28": 0.0,
          "29": 0.0,
          "30": 0.0,
          "31": 0.0
        }
      }
    },
    "native-I4": {
      "engine": "native",
      "candidate": "I4",
      "records": 4714,
      "failed_requests": [],
      "final_document_expected": 128,
      "final_document_present": 128,
      "final_document_failed": 0,
      "quality_complete": true,
      "nll": 1.8817406798660223,
      "ppl": 6.564922350369657,
      "scored_tokens": 130944,
      "domains": {
        "ko": {
          "documents": 64,
          "nll": 2.1255431679640813,
          "ppl": 8.377446613167
        },
        "en": {
          "documents": 32,
          "nll": 2.196320837668671,
          "ppl": 8.991870017239366
        },
        "code": {
          "documents": 32,
          "nll": 1.079555545867256,
          "ppl": 2.94337106687159
        }
      },
      "delta_nll": {
        "delta": 0.030243037076852775,
        "ci95": [
          0.027901765067122962,
          0.03277231822680255
        ],
        "resampling_unit": "paired document",
        "replicates": 2000
      },
      "ppl_ratio_vs_bf16": {
        "ratio": 1.0307050030466838,
        "ci95": [
          1.028294665002294,
          1.0333152454089167
        ]
      },
      "document_delta_tail": {
        "mean": 0.030243037076852775,
        "median": 0.027785089374209863,
        "p95": 0.05720809367587929,
        "p99": 0.07025643392598449,
        "max": 0.07380433181319565
      },
      "worst20_documents": [
        {
          "id": "final-049",
          "domain": "code",
          "delta_nll": 0.07380433181319565,
          "nll": 0.597948293947374
        },
        {
          "id": "final-003",
          "domain": "ko",
          "delta_nll": 0.07043613565274903,
          "nll": 1.4895878737949164
        },
        {
          "id": "final-012",
          "domain": "code",
          "delta_nll": 0.06977057370176921,
          "nll": 0.34462881442939397
        },
        {
          "id": "final-001",
          "domain": "code",
          "delta_nll": 0.063903892333006,
          "nll": 0.26384155171442636
        },
        {
          "id": "final-046",
          "domain": "code",
          "delta_nll": 0.06124876146265229,
          "nll": 1.3369401815437045
        },
        {
          "id": "final-126",
          "domain": "ko",
          "delta_nll": 0.059376084507181084,
          "nll": 2.4266440783292293
        },
        {
          "id": "final-075",
          "domain": "ko",
          "delta_nll": 0.05821306199108402,
          "nll": 1.9939955712261537
        },
        {
          "id": "final-121",
          "domain": "code",
          "delta_nll": 0.05534172394764201,
          "nll": 1.159221655349253
        },
        {
          "id": "final-087",
          "domain": "en",
          "delta_nll": 0.053838557034368656,
          "nll": 2.007992978732109
        },
        {
          "id": "final-073",
          "domain": "en",
          "delta_nll": 0.05356137828003158,
          "nll": 2.2537412219693067
        },
        {
          "id": "final-077",
          "domain": "ko",
          "delta_nll": 0.05005930666352927,
          "nll": 2.2500460776789875
        },
        {
          "id": "final-023",
          "domain": "code",
          "delta_nll": 0.048608831579902,
          "nll": 0.43438310067466784
        },
        {
          "id": "final-090",
          "domain": "code",
          "delta_nll": 0.047893375054855136,
          "nll": 0.4292391852312617
        },
        {
          "id": "final-067",
          "domain": "code",
          "delta_nll": 0.04773035563201086,
          "nll": 0.30584377546058883
        },
        {
          "id": "final-024",
          "domain": "ko",
          "delta_nll": 0.04722641881378296,
          "nll": 2.178383172753178
        },
        {
          "id": "final-007",
          "domain": "ko",
          "delta_nll": 0.045400679665202066,
          "nll": 1.9866574135169721
        },
        {
          "id": "final-035",
          "domain": "ko",
          "delta_nll": 0.04402014692743672,
          "nll": 2.133459971062297
        },
        {
          "id": "final-109",
          "domain": "en",
          "delta_nll": 0.04381307316671137,
          "nll": 2.139233080749062
        },
        {
          "id": "final-031",
          "domain": "ko",
          "delta_nll": 0.043743082225102155,
          "nll": 1.9211921424129628
        },
        {
          "id": "final-118",
          "domain": "ko",
          "delta_nll": 0.043503716774552975,
          "nll": 2.1575480676744987
        }
      ],
      "token_delta_tail": {
        "mean": 0.030243037076852657,
        "median": 0.0004181608805549075,
        "p95": 0.4711779953857906,
        "p99": 0.9814744929429333,
        "max": 9.69035419171897
      },
      "sampled_full_vocab_kl": {
        "mean": 0.03342629550417496,
        "median": 0.013606945214072328,
        "p95": 0.10202216382966092,
        "p99": 0.32428529985428034,
        "max": 10.737569786955627
      },
      "kl_positions": 4096,
      "sampled_top1_agreement": 0.9345703125,
      "tasks": {
        "kmmlu": {
          "n": 500,
          "accuracy": 0.478,
          "failures": 0,
          "subject_scores": {
            "Accounting": {
              "n": 12,
              "accuracy": 0.5
            },
            "Agricultural-Sciences": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Aviation-Engineering-and-Maintenance": {
              "n": 12,
              "accuracy": 0.5
            },
            "Biology": {
              "n": 12,
              "accuracy": 0.25
            },
            "Chemical-Engineering": {
              "n": 12,
              "accuracy": 0.25
            },
            "Chemistry": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Civil-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Computer-Science": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Construction": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Criminal-Law": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Ecology": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Economics": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Education": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Electrical-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Electronics-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Energy-Management": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Environmental-Science": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Fashion": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Food-Processing": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Gas-Technology-and-Engineering": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Geomatics": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Health": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Industrial-Engineer": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Information-Technology": {
              "n": 11,
              "accuracy": 1.0
            },
            "Interior-Architecture-and-Design": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Korean-History": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Law": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Machine-Design-and-Manufacturing": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Management": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Maritime-Engineering": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Marketing": {
              "n": 11,
              "accuracy": 1.0
            },
            "Materials-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Math": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Mechanical-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Nondestructive-Testing": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Patent": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Political-Science-and-Sociology": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Psychology": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Public-Safety": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Railway-and-Automotive-Engineering": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Real-Estate": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Refrigerating-Machinery": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Social-Welfare": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Taxation": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Telecommunications-and-Wireless-Technology": {
              "n": 11,
              "accuracy": 0.6363636363636364
            }
          },
          "macro_accuracy": 0.47895622895622886,
          "delta_accuracy": {
            "delta": -0.004,
            "ci95": [
              -0.03,
              0.022
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "kmmlu-Accounting-96",
            "kmmlu-Agricultural-Sciences-226",
            "kmmlu-Biology-379",
            "kmmlu-Chemical-Engineering-811",
            "kmmlu-Chemical-Engineering-584",
            "kmmlu-Chemistry-309",
            "kmmlu-Electrical-Engineering-481",
            "kmmlu-Electronics-Engineering-810",
            "kmmlu-Electronics-Engineering-83",
            "kmmlu-Energy-Management-645",
            "kmmlu-Energy-Management-431",
            "kmmlu-Environmental-Science-342",
            "kmmlu-Interior-Architecture-and-Design-849",
            "kmmlu-Management-192",
            "kmmlu-Management-22",
            "kmmlu-Math-208",
            "kmmlu-Nondestructive-Testing-286",
            "kmmlu-Patent-99",
            "kmmlu-Political-Science-and-Sociology-76",
            "kmmlu-Political-Science-and-Sociology-123",
            "kmmlu-Refrigerating-Machinery-247",
            "kmmlu-Social-Welfare-952"
          ],
          "improvements": [
            "kmmlu-Chemistry-130",
            "kmmlu-Economics-26",
            "kmmlu-Electrical-Engineering-112",
            "kmmlu-Food-Processing-919",
            "kmmlu-Geomatics-951",
            "kmmlu-Geomatics-272",
            "kmmlu-Geomatics-900",
            "kmmlu-Interior-Architecture-and-Design-723",
            "kmmlu-Interior-Architecture-and-Design-561",
            "kmmlu-Korean-History-89",
            "kmmlu-Maritime-Engineering-234",
            "kmmlu-Marketing-80",
            "kmmlu-Materials-Engineering-321",
            "kmmlu-Patent-48",
            "kmmlu-Psychology-892",
            "kmmlu-Railway-and-Automotive-Engineering-834",
            "kmmlu-Railway-and-Automotive-Engineering-177",
            "kmmlu-Real-Estate-155",
            "kmmlu-Social-Welfare-606",
            "kmmlu-Social-Welfare-706"
          ]
        },
        "arc": {
          "n": 500,
          "accuracy": 0.524,
          "failures": 0,
          "accuracy_norm": 0.554,
          "delta_accuracy": {
            "delta": 0.006,
            "ci95": [
              -0.014,
              0.026
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "arc-TIMSS_2003_8_pg87",
            "arc-Mercury_7268328",
            "arc-ACTAAP_2010_7_12",
            "arc-Mercury_SC_407689",
            "arc-Mercury_7211033",
            "arc-Mercury_SC_415453",
            "arc-Mercury_SC_401366",
            "arc-Mercury_7108990",
            "arc-MCAS_1999_4_23",
            "arc-NYSEDREGENTS_2008_8_27",
            "arc-Mercury_7212328",
            "arc-Mercury_7201740"
          ],
          "improvements": [
            "arc-Mercury_7143308",
            "arc-Mercury_SC_400594",
            "arc-MEAP_2005_8_39",
            "arc-Mercury_183033",
            "arc-TIMSS_2007_8_pg60",
            "arc-Mercury_189053",
            "arc-NYSEDREGENTS_2008_4_15",
            "arc-NYSEDREGENTS_2012_4_29",
            "arc-Mercury_SC_409241",
            "arc-OHAT_2007_5_15",
            "arc-Mercury_7216248",
            "arc-Mercury_189070",
            "arc-NCEOGA_2013_8_28",
            "arc-Mercury_411732",
            "arc-Mercury_7210245"
          ],
          "delta_accuracy_norm": {
            "delta": -0.008,
            "ci95": [
              -0.026,
              0.01
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          }
        },
        "ifeval": {
          "n": 541,
          "accuracy": 0.8188539741219963,
          "failures": 0,
          "instruction_strict": 0.8717026378896883,
          "truncated": 13,
          "delta_accuracy": {
            "delta": -0.0073937153419593345,
            "ci95": [
              -0.03512014787430684,
              0.02033271719038817
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "ifeval-1075",
            "ifeval-1129",
            "ifeval-1130",
            "ifeval-1379",
            "ifeval-1516",
            "ifeval-1566",
            "ifeval-1609",
            "ifeval-1670",
            "ifeval-1825",
            "ifeval-1996",
            "ifeval-2034",
            "ifeval-2035",
            "ifeval-2416",
            "ifeval-2439",
            "ifeval-2497",
            "ifeval-2605",
            "ifeval-2844",
            "ifeval-2880",
            "ifeval-2912",
            "ifeval-2969",
            "ifeval-301",
            "ifeval-3048",
            "ifeval-3069",
            "ifeval-3098",
            "ifeval-3243",
            "ifeval-3326",
            "ifeval-3351",
            "ifeval-3362",
            "ifeval-3428",
            "ifeval-3455",
            "ifeval-371"
          ],
          "improvements": [
            "ifeval-1021",
            "ifeval-1092",
            "ifeval-1219",
            "ifeval-1265",
            "ifeval-1348",
            "ifeval-1372",
            "ifeval-1389",
            "ifeval-1582",
            "ifeval-1629",
            "ifeval-1645",
            "ifeval-2230",
            "ifeval-2266",
            "ifeval-2273",
            "ifeval-2299",
            "ifeval-2386",
            "ifeval-2534",
            "ifeval-2571",
            "ifeval-2591",
            "ifeval-2616",
            "ifeval-2662",
            "ifeval-2674",
            "ifeval-2736",
            "ifeval-2929",
            "ifeval-3035",
            "ifeval-3114",
            "ifeval-3329",
            "ifeval-3409"
          ]
        }
      },
      "speed": {
        "512": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.0321546,
            "min": 0.0318315,
            "max": 0.0434728
          },
          "prefill_seconds": {
            "median": 0.0321327,
            "min": 0.0318092,
            "max": 0.0434489
          },
          "decode_seconds": {
            "median": 0.8676278,
            "min": 0.8527025,
            "max": 0.9072525
          },
          "seconds": {
            "median": 0.8995087,
            "min": 0.8845344,
            "max": 0.9394076
          },
          "prefill_tokens_per_second": {
            "median": 15933.924008875694,
            "min": 11783.957706639294,
            "max": 16095.972234447894
          },
          "decode_tokens_per_second": {
            "median": 293.90482877565705,
            "min": 281.0683905527954,
            "max": 299.04919945702045
          },
          "peak_device_vram_bytes": 7447752704
        },
        "4096": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.2543446,
            "min": 0.2522706,
            "max": 0.3004729
          },
          "prefill_seconds": {
            "median": 0.254317,
            "min": 0.2522488,
            "max": 0.3004501
          },
          "decode_seconds": {
            "median": 0.8634475,
            "min": 0.8604546,
            "max": 0.9443029
          },
          "seconds": {
            "median": 1.1258522,
            "min": 1.1138621,
            "max": 1.1965739
          },
          "prefill_tokens_per_second": {
            "median": 16105.88360196133,
            "min": 13632.879469835423,
            "max": 16237.936513473998
          },
          "decode_tokens_per_second": {
            "median": 295.32774140871334,
            "min": 270.04047112425474,
            "max": 296.3549732897006
          },
          "peak_device_vram_bytes": 7447752704
        }
      },
      "speed_source": "dedicated serial benchmark",
      "runtime_memory_breakdown": {
        "source": "llama.cpp CUDA0 buffer log",
        "model_mib": 3029.92,
        "kv_mib": 256.0,
        "recurrent_state_mib": 50.25,
        "compute_mib": 495.0
      },
      "block_diagnostics": {
        "scope": "six fixed validation documents, 256 tokens, cumulative error",
        "layer_mean_relative_l2": {
          "0": 0.05945174334560872,
          "1": 0.0783909388325922,
          "2": 0.11083954845195043,
          "3": 0.10881876938408984,
          "4": 0.12105555181836157,
          "5": 0.1248304704592816,
          "6": 0.12745386693640845,
          "7": 0.13842950514868071,
          "8": 0.14042917735344837,
          "9": 0.14341035886328066,
          "10": 0.1416100387087346,
          "11": 0.1514952923869246,
          "12": 0.1542053223190083,
          "13": 0.15443041750249387,
          "14": 0.15634749036682782,
          "15": 0.16304742884408438,
          "16": 0.16570423427182082,
          "17": 0.16057487752414987,
          "18": 0.15351608660486268,
          "19": 0.16190285638406396,
          "20": 0.15751221313613437,
          "21": 0.1564099730934164,
          "22": 0.1591623240098972,
          "23": 0.15705860661558005,
          "24": 0.1569009512306538,
          "25": 0.1588038424931714,
          "26": 0.17036119028647567,
          "27": 0.1690551457664339,
          "28": 0.16344070332987504,
          "29": 0.16843632358156227,
          "30": 0.17760223034667236,
          "31": 0.1767388234421695
        }
      }
    },
    "native-K0": {
      "engine": "native",
      "candidate": "K0",
      "records": 4714,
      "failed_requests": [],
      "final_document_expected": 128,
      "final_document_present": 128,
      "final_document_failed": 0,
      "quality_complete": true,
      "nll": 1.8860503484055466,
      "ppl": 6.593276043311702,
      "scored_tokens": 130944,
      "domains": {
        "ko": {
          "documents": 64,
          "nll": 2.1341204545084773,
          "ppl": 8.449611419962743
        },
        "en": {
          "documents": 32,
          "nll": 2.1928519689817065,
          "ppl": 8.960732438209085
        },
        "code": {
          "documents": 32,
          "nll": 1.0831085156235258,
          "ppl": 2.953847375236333
        }
      },
      "delta_nll": {
        "delta": 0.03455270561637727,
        "ci95": [
          0.03165340037672631,
          0.0372973156617618
        ],
        "resampling_unit": "paired document",
        "replicates": 2000
      },
      "ppl_ratio_vs_bf16": {
        "ratio": 1.0351565855042542,
        "ci95": [
          1.0321596971384261,
          1.0380015890941678
        ]
      },
      "document_delta_tail": {
        "mean": 0.03455270561637727,
        "median": 0.03247442800900782,
        "p95": 0.06356939068752121,
        "p99": 0.07378202164310624,
        "max": 0.08018221477548558
      },
      "worst20_documents": [
        {
          "id": "final-049",
          "domain": "code",
          "delta_nll": 0.08018221477548558,
          "nll": 0.604326176909664
        },
        {
          "id": "final-063",
          "domain": "ko",
          "delta_nll": 0.07451020032595729,
          "nll": 2.004895962536779
        },
        {
          "id": "final-012",
          "domain": "code",
          "delta_nll": 0.07181324224132374,
          "nll": 0.3466714829689485
        },
        {
          "id": "final-126",
          "domain": "ko",
          "delta_nll": 0.0702623463149742,
          "nll": 2.4375303401370223
        },
        {
          "id": "final-124",
          "domain": "ko",
          "delta_nll": 0.0671015043645476,
          "nll": 2.1144498761124653
        },
        {
          "id": "final-009",
          "domain": "ko",
          "delta_nll": 0.06675581638980901,
          "nll": 2.302815680099088
        },
        {
          "id": "final-023",
          "domain": "code",
          "delta_nll": 0.06359360603544861,
          "nll": 0.4493678751302145
        },
        {
          "id": "final-089",
          "domain": "ko",
          "delta_nll": 0.06352441932708461,
          "nll": 2.7434548353146533
        },
        {
          "id": "final-035",
          "domain": "ko",
          "delta_nll": 0.06284441766894783,
          "nll": 2.1522842418038084
        },
        {
          "id": "final-001",
          "domain": "code",
          "delta_nll": 0.06176999885005194,
          "nll": 0.26170765823147224
        },
        {
          "id": "final-034",
          "domain": "ko",
          "delta_nll": 0.06131229495963942,
          "nll": 2.0421069230633924
        },
        {
          "id": "final-103",
          "domain": "en",
          "delta_nll": 0.06106495530604327,
          "nll": 2.13662150953151
        },
        {
          "id": "final-104",
          "domain": "ko",
          "delta_nll": 0.05810418220928784,
          "nll": 2.661881415109549
        },
        {
          "id": "final-081",
          "domain": "code",
          "delta_nll": 0.05599665249793446,
          "nll": 0.4963711207885065
        },
        {
          "id": "final-051",
          "domain": "ko",
          "delta_nll": 0.05571570246230585,
          "nll": 2.746760436051575
        },
        {
          "id": "final-110",
          "domain": "ko",
          "delta_nll": 0.0550843839661047,
          "nll": 2.3278257452490507
        },
        {
          "id": "final-085",
          "domain": "ko",
          "delta_nll": 0.05464285702737248,
          "nll": 2.6868628290525396
        },
        {
          "id": "final-075",
          "domain": "ko",
          "delta_nll": 0.05126414635944629,
          "nll": 1.9870466555945159
        },
        {
          "id": "final-046",
          "domain": "code",
          "delta_nll": 0.050723821104224756,
          "nll": 1.326415241185277
        },
        {
          "id": "final-003",
          "domain": "ko",
          "delta_nll": 0.049201561411098535,
          "nll": 1.4683532995532658
        }
      ],
      "token_delta_tail": {
        "mean": 0.03455270561637709,
        "median": 0.0007564102433814668,
        "p95": 0.5346366993519095,
        "p99": 1.1218961794474132,
        "max": 16.885432344463545
      },
      "sampled_full_vocab_kl": {
        "mean": 0.04303441856732754,
        "median": 0.017307583372364023,
        "p95": 0.13528341212096812,
        "p99": 0.439438270231327,
        "max": 3.120973678948202
      },
      "kl_positions": 4096,
      "sampled_top1_agreement": 0.9169921875,
      "tasks": {
        "kmmlu": {
          "n": 500,
          "accuracy": 0.494,
          "failures": 0,
          "subject_scores": {
            "Accounting": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Agricultural-Sciences": {
              "n": 12,
              "accuracy": 0.5
            },
            "Aviation-Engineering-and-Maintenance": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Biology": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Chemical-Engineering": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Chemistry": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Civil-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Computer-Science": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Construction": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Criminal-Law": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Ecology": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Economics": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Education": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Electrical-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Electronics-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Energy-Management": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Environmental-Science": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Fashion": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Food-Processing": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Gas-Technology-and-Engineering": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Geomatics": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Health": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Industrial-Engineer": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Information-Technology": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Interior-Architecture-and-Design": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Korean-History": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Law": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Machine-Design-and-Manufacturing": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Management": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Maritime-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Marketing": {
              "n": 11,
              "accuracy": 1.0
            },
            "Materials-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Math": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Mechanical-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Nondestructive-Testing": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Patent": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Political-Science-and-Sociology": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Psychology": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Public-Safety": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Railway-and-Automotive-Engineering": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Real-Estate": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Refrigerating-Machinery": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Social-Welfare": {
              "n": 11,
              "accuracy": 0.9090909090909091
            },
            "Taxation": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Telecommunications-and-Wireless-Technology": {
              "n": 11,
              "accuracy": 0.7272727272727273
            }
          },
          "macro_accuracy": 0.49461279461279467,
          "delta_accuracy": {
            "delta": 0.012,
            "ci95": [
              -0.02,
              0.044
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "kmmlu-Accounting-52",
            "kmmlu-Accounting-96",
            "kmmlu-Aviation-Engineering-and-Maintenance-899",
            "kmmlu-Biology-379",
            "kmmlu-Chemical-Engineering-584",
            "kmmlu-Criminal-Law-73",
            "kmmlu-Ecology-998",
            "kmmlu-Economics-104",
            "kmmlu-Education-20",
            "kmmlu-Electrical-Engineering-481",
            "kmmlu-Electronics-Engineering-294",
            "kmmlu-Electronics-Engineering-83",
            "kmmlu-Energy-Management-645",
            "kmmlu-Energy-Management-155",
            "kmmlu-Environmental-Science-438",
            "kmmlu-Environmental-Science-342",
            "kmmlu-Fashion-419",
            "kmmlu-Food-Processing-148",
            "kmmlu-Information-Technology-977",
            "kmmlu-Information-Technology-635",
            "kmmlu-Interior-Architecture-and-Design-849",
            "kmmlu-Machine-Design-and-Manufacturing-699",
            "kmmlu-Mechanical-Engineering-305",
            "kmmlu-Nondestructive-Testing-286",
            "kmmlu-Patent-28",
            "kmmlu-Patent-99",
            "kmmlu-Political-Science-and-Sociology-123",
            "kmmlu-Refrigerating-Machinery-797",
            "kmmlu-Social-Welfare-952"
          ],
          "improvements": [
            "kmmlu-Biology-596",
            "kmmlu-Biology-731",
            "kmmlu-Chemical-Engineering-970",
            "kmmlu-Construction-396",
            "kmmlu-Criminal-Law-172",
            "kmmlu-Economics-74",
            "kmmlu-Electrical-Engineering-962",
            "kmmlu-Environmental-Science-835",
            "kmmlu-Fashion-767",
            "kmmlu-Food-Processing-85",
            "kmmlu-Food-Processing-919",
            "kmmlu-Geomatics-900",
            "kmmlu-Health-58",
            "kmmlu-Industrial-Engineer-141",
            "kmmlu-Industrial-Engineer-927",
            "kmmlu-Interior-Architecture-and-Design-723",
            "kmmlu-Interior-Architecture-and-Design-561",
            "kmmlu-Korean-History-64",
            "kmmlu-Korean-History-53",
            "kmmlu-Law-11",
            "kmmlu-Law-694",
            "kmmlu-Machine-Design-and-Manufacturing-449",
            "kmmlu-Marketing-80",
            "kmmlu-Math-68",
            "kmmlu-Nondestructive-Testing-312",
            "kmmlu-Patent-48",
            "kmmlu-Public-Safety-152",
            "kmmlu-Railway-and-Automotive-Engineering-834",
            "kmmlu-Railway-and-Automotive-Engineering-177",
            "kmmlu-Real-Estate-172",
            "kmmlu-Real-Estate-155",
            "kmmlu-Social-Welfare-606",
            "kmmlu-Social-Welfare-706",
            "kmmlu-Social-Welfare-19",
            "kmmlu-Telecommunications-and-Wireless-Technology-713"
          ]
        },
        "arc": {
          "n": 500,
          "accuracy": 0.498,
          "failures": 0,
          "accuracy_norm": 0.53,
          "delta_accuracy": {
            "delta": -0.02,
            "ci95": [
              -0.04205,
              0.002
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "arc-Mercury_7025060",
            "arc-Mercury_7268328",
            "arc-Mercury_7211680",
            "arc-MCAS_2004_8_14",
            "arc-Mercury_7228725",
            "arc-Mercury_7007420",
            "arc-Mercury_7093048",
            "arc-ACTAAP_2010_7_12",
            "arc-Mercury_7223493",
            "arc-Mercury_SC_402257",
            "arc-Mercury_SC_407689",
            "arc-Mercury_7042945",
            "arc-Mercury_SC_414156",
            "arc-MDSA_2007_8_42",
            "arc-Mercury_7212328",
            "arc-Mercury_7201740",
            "arc-Mercury_409295",
            "arc-Mercury_SC_408321",
            "arc-Mercury_402364",
            "arc-Mercury_416536",
            "arc-Mercury_7215670"
          ],
          "improvements": [
            "arc-Mercury_SC_405885",
            "arc-TIMSS_1995_8_K18",
            "arc-NYSEDREGENTS_2008_4_15",
            "arc-Mercury_SC_409241",
            "arc-Mercury_SC_402164",
            "arc-Mercury_7283833",
            "arc-Mercury_404132",
            "arc-MEA_2016_8_15",
            "arc-Mercury_415303",
            "arc-Mercury_189070",
            "arc-Mercury_7210245"
          ],
          "delta_accuracy_norm": {
            "delta": -0.032,
            "ci95": [
              -0.058,
              -0.01
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          }
        },
        "ifeval": {
          "n": 541,
          "accuracy": 0.8040665434380776,
          "failures": 0,
          "instruction_strict": 0.8609112709832134,
          "truncated": 18,
          "delta_accuracy": {
            "delta": -0.022181146025878003,
            "ci95": [
              -0.04990757855822551,
              0.005545286506469501
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "ifeval-1203",
            "ifeval-1379",
            "ifeval-1392",
            "ifeval-1601",
            "ifeval-1644",
            "ifeval-1670",
            "ifeval-1713",
            "ifeval-179",
            "ifeval-1823",
            "ifeval-1922",
            "ifeval-1967",
            "ifeval-1996",
            "ifeval-2034",
            "ifeval-2035",
            "ifeval-2069",
            "ifeval-2084",
            "ifeval-2192",
            "ifeval-2275",
            "ifeval-2589",
            "ifeval-2605",
            "ifeval-2811",
            "ifeval-2844",
            "ifeval-2912",
            "ifeval-2918",
            "ifeval-2943",
            "ifeval-301",
            "ifeval-3048",
            "ifeval-3098",
            "ifeval-3109",
            "ifeval-3245",
            "ifeval-3287",
            "ifeval-3351",
            "ifeval-3362",
            "ifeval-3429",
            "ifeval-3479",
            "ifeval-3703",
            "ifeval-3724",
            "ifeval-3748"
          ],
          "improvements": [
            "ifeval-1021",
            "ifeval-1132",
            "ifeval-1265",
            "ifeval-127",
            "ifeval-1348",
            "ifeval-1372",
            "ifeval-1389",
            "ifeval-1580",
            "ifeval-1582",
            "ifeval-1629",
            "ifeval-1944",
            "ifeval-2078",
            "ifeval-2266",
            "ifeval-2299",
            "ifeval-2386",
            "ifeval-2616",
            "ifeval-2674",
            "ifeval-2736",
            "ifeval-2749",
            "ifeval-2929",
            "ifeval-3035",
            "ifeval-3089",
            "ifeval-3114",
            "ifeval-3409",
            "ifeval-3617",
            "ifeval-3691"
          ]
        }
      },
      "speed": {
        "512": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.0329756,
            "min": 0.0329278,
            "max": 0.0331611
          },
          "prefill_seconds": {
            "median": 0.0329557,
            "min": 0.0329036,
            "max": 0.0331387
          },
          "decode_seconds": {
            "median": 0.892806,
            "min": 0.8877685,
            "max": 0.9660976
          },
          "seconds": {
            "median": 0.9257346,
            "min": 0.9208202,
            "max": 0.9990738
          },
          "prefill_tokens_per_second": {
            "median": 15536.007428153553,
            "min": 15450.213798368675,
            "max": 15560.60734995563
          },
          "decode_tokens_per_second": {
            "median": 285.6163601051068,
            "min": 263.9484871921843,
            "max": 287.2370443420779
          },
          "peak_device_vram_bytes": 7613911040
        },
        "4096": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.2622093,
            "min": 0.2606742,
            "max": 0.2656708
          },
          "prefill_seconds": {
            "median": 0.2621853,
            "min": 0.2606478,
            "max": 0.2656398
          },
          "decode_seconds": {
            "median": 0.9130032,
            "min": 0.8931073,
            "max": 0.9816207
          },
          "seconds": {
            "median": 1.1786744,
            "min": 1.1552752,
            "max": 1.2436598
          },
          "prefill_tokens_per_second": {
            "median": 15622.538715938688,
            "min": 15419.376162758745,
            "max": 15714.69239333691
          },
          "decode_tokens_per_second": {
            "median": 279.2980353190438,
            "min": 259.77447297107733,
            "max": 285.520004147318
          },
          "peak_device_vram_bytes": 7613911040
        }
      },
      "speed_source": "dedicated serial benchmark",
      "runtime_memory_breakdown": {
        "source": "llama.cpp CUDA0 buffer log",
        "model_mib": 3136.17,
        "kv_mib": 256.0,
        "recurrent_state_mib": 50.25,
        "compute_mib": 495.0
      },
      "block_diagnostics": {
        "scope": "six fixed validation documents, 256 tokens, cumulative error",
        "layer_mean_relative_l2": {
          "0": 0.1080037999628915,
          "1": 0.1252887376234116,
          "2": 0.1620984789751663,
          "3": 0.15337773330650703,
          "4": 0.1616344719309899,
          "5": 0.1611653406666588,
          "6": 0.16217424258670984,
          "7": 0.17182376851528638,
          "8": 0.17156664729839063,
          "9": 0.17347707465485773,
          "10": 0.17056636553567153,
          "11": 0.18038459705435014,
          "12": 0.1833024384539849,
          "13": 0.182970077747693,
          "14": 0.1850569469848723,
          "15": 0.19286805654220993,
          "16": 0.19658882739647424,
          "17": 0.18980855088131157,
          "18": 0.18102348399361357,
          "19": 0.18936113912570404,
          "20": 0.18343657796264012,
          "21": 0.18077561092237135,
          "22": 0.18159410458906544,
          "23": 0.18179332116637817,
          "24": 0.1815043195036944,
          "25": 0.18360318369987208,
          "26": 0.19173765022077158,
          "27": 0.1967464658714716,
          "28": 0.19071340262263303,
          "29": 0.19463557370662488,
          "30": 0.20563378247306577,
          "31": 0.20309930844598156
        }
      }
    },
    "native-K1": {
      "engine": "native",
      "candidate": "K1",
      "records": 4714,
      "failed_requests": [],
      "final_document_expected": 128,
      "final_document_present": 128,
      "final_document_failed": 0,
      "quality_complete": true,
      "nll": 1.8744917661937208,
      "ppl": 6.517505861634886,
      "scored_tokens": 130944,
      "domains": {
        "ko": {
          "documents": 64,
          "nll": 2.120660871750281,
          "ppl": 8.33664512082013
        },
        "en": {
          "documents": 32,
          "nll": 2.185382178000655,
          "ppl": 8.894047013026917
        },
        "code": {
          "documents": 32,
          "nll": 1.0712631432736677,
          "ppl": 2.919064368751241
        }
      },
      "delta_nll": {
        "delta": 0.0229941234045515,
        "ci95": [
          0.020681926587697443,
          0.02522469552367121
        ],
        "resampling_unit": "paired document",
        "replicates": 2000
      },
      "ppl_ratio_vs_bf16": {
        "ratio": 1.0232605262413974,
        "ci95": [
          1.0208972797082416,
          1.0255455301273968
        ]
      },
      "document_delta_tail": {
        "mean": 0.0229941234045515,
        "median": 0.02252778320885787,
        "p95": 0.04751316356663568,
        "p99": 0.054710523856905405,
        "max": 0.06654257498736983
      },
      "worst20_documents": [
        {
          "id": "final-049",
          "domain": "code",
          "delta_nll": 0.06654257498736983,
          "nll": 0.5906865371215482
        },
        {
          "id": "final-012",
          "domain": "code",
          "delta_nll": 0.05634159304970733,
          "nll": 0.3311998337773321
        },
        {
          "id": "final-009",
          "domain": "ko",
          "delta_nll": 0.05030059603932973,
          "nll": 2.2863604597486087
        },
        {
          "id": "final-126",
          "domain": "ko",
          "delta_nll": 0.04956280792329472,
          "nll": 2.4168308017453426
        },
        {
          "id": "final-046",
          "domain": "code",
          "delta_nll": 0.048418691271459625,
          "nll": 1.3241101113525118
        },
        {
          "id": "final-015",
          "domain": "ko",
          "delta_nll": 0.048408062264993355,
          "nll": 1.264200872124446
        },
        {
          "id": "final-003",
          "domain": "ko",
          "delta_nll": 0.04811811956815295,
          "nll": 1.4672698577103203
        },
        {
          "id": "final-034",
          "domain": "ko",
          "delta_nll": 0.04638967384953223,
          "nll": 2.027184301953285
        },
        {
          "id": "final-124",
          "domain": "ko",
          "delta_nll": 0.046259074937886495,
          "nll": 2.0936074466858043
        },
        {
          "id": "final-022",
          "domain": "ko",
          "delta_nll": 0.046106623765750976,
          "nll": 2.1564485687087753
        },
        {
          "id": "final-103",
          "domain": "en",
          "delta_nll": 0.043579037724630364,
          "nll": 2.1191355919500974
        },
        {
          "id": "final-063",
          "domain": "ko",
          "delta_nll": 0.04177426178620086,
          "nll": 1.9721600239970225
        },
        {
          "id": "final-062",
          "domain": "ko",
          "delta_nll": 0.040493611630352006,
          "nll": 2.173768449703913
        },
        {
          "id": "final-014",
          "domain": "ko",
          "delta_nll": 0.04045411986351544,
          "nll": 2.6759600125208025
        },
        {
          "id": "final-081",
          "domain": "code",
          "delta_nll": 0.03840746082056484,
          "nll": 0.4787819291111369
        },
        {
          "id": "final-118",
          "domain": "ko",
          "delta_nll": 0.03802793493424073,
          "nll": 2.152072285834186
        },
        {
          "id": "final-033",
          "domain": "ko",
          "delta_nll": 0.03770490404312685,
          "nll": 2.3835211978340647
        },
        {
          "id": "final-024",
          "domain": "ko",
          "delta_nll": 0.03696994464438918,
          "nll": 2.1681266985837846
        },
        {
          "id": "final-089",
          "domain": "ko",
          "delta_nll": 0.03643101492258621,
          "nll": 2.7163614309101547
        },
        {
          "id": "final-090",
          "domain": "code",
          "delta_nll": 0.035225503689578906,
          "nll": 0.41657131386598545
        }
      ],
      "token_delta_tail": {
        "mean": 0.022994123404551273,
        "median": 0.0005994020219404206,
        "p95": 0.4196738743059049,
        "p99": 0.8965297193657,
        "max": 6.741763657833058
      },
      "sampled_full_vocab_kl": {
        "mean": 0.028562542985934247,
        "median": 0.011532071488087464,
        "p95": 0.08724382438002827,
        "p99": 0.2767615327215544,
        "max": 2.6137904837426174
      },
      "kl_positions": 4096,
      "sampled_top1_agreement": 0.931884765625,
      "tasks": {
        "kmmlu": {
          "n": 500,
          "accuracy": 0.466,
          "failures": 0,
          "subject_scores": {
            "Accounting": {
              "n": 12,
              "accuracy": 0.5833333333333334
            },
            "Agricultural-Sciences": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Aviation-Engineering-and-Maintenance": {
              "n": 12,
              "accuracy": 0.5833333333333334
            },
            "Biology": {
              "n": 12,
              "accuracy": 0.5
            },
            "Chemical-Engineering": {
              "n": 12,
              "accuracy": 0.3333333333333333
            },
            "Chemistry": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Civil-Engineering": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Computer-Science": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Construction": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Criminal-Law": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Ecology": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Economics": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Education": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Electrical-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Electronics-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Energy-Management": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Environmental-Science": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Fashion": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Food-Processing": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Gas-Technology-and-Engineering": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Geomatics": {
              "n": 11,
              "accuracy": 0.0
            },
            "Health": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Industrial-Engineer": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Information-Technology": {
              "n": 11,
              "accuracy": 0.9090909090909091
            },
            "Interior-Architecture-and-Design": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Korean-History": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Law": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Machine-Design-and-Manufacturing": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Management": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Maritime-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Marketing": {
              "n": 11,
              "accuracy": 0.9090909090909091
            },
            "Materials-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Math": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Mechanical-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Nondestructive-Testing": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Patent": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Political-Science-and-Sociology": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Psychology": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Public-Safety": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Railway-and-Automotive-Engineering": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Real-Estate": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Refrigerating-Machinery": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Social-Welfare": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Taxation": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Telecommunications-and-Wireless-Technology": {
              "n": 11,
              "accuracy": 0.6363636363636364
            }
          },
          "macro_accuracy": 0.4658249158249158,
          "delta_accuracy": {
            "delta": -0.016,
            "ci95": [
              -0.042,
              0.008
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "kmmlu-Agricultural-Sciences-226",
            "kmmlu-Chemical-Engineering-811",
            "kmmlu-Chemistry-486",
            "kmmlu-Civil-Engineering-837",
            "kmmlu-Ecology-95",
            "kmmlu-Economics-129",
            "kmmlu-Electronics-Engineering-810",
            "kmmlu-Electronics-Engineering-83",
            "kmmlu-Energy-Management-645",
            "kmmlu-Geomatics-629",
            "kmmlu-Industrial-Engineer-130",
            "kmmlu-Industrial-Engineer-706",
            "kmmlu-Information-Technology-635",
            "kmmlu-Interior-Architecture-and-Design-849",
            "kmmlu-Interior-Architecture-and-Design-560",
            "kmmlu-Machine-Design-and-Manufacturing-345",
            "kmmlu-Machine-Design-and-Manufacturing-699",
            "kmmlu-Maritime-Engineering-413",
            "kmmlu-Nondestructive-Testing-286",
            "kmmlu-Patent-99",
            "kmmlu-Political-Science-and-Sociology-76",
            "kmmlu-Public-Safety-176",
            "kmmlu-Refrigerating-Machinery-797",
            "kmmlu-Refrigerating-Machinery-247",
            "kmmlu-Social-Welfare-952"
          ],
          "improvements": [
            "kmmlu-Aviation-Engineering-and-Maintenance-216",
            "kmmlu-Biology-596",
            "kmmlu-Biology-731",
            "kmmlu-Chemistry-130",
            "kmmlu-Criminal-Law-172",
            "kmmlu-Economics-26",
            "kmmlu-Energy-Management-826",
            "kmmlu-Food-Processing-919",
            "kmmlu-Industrial-Engineer-927",
            "kmmlu-Interior-Architecture-and-Design-947",
            "kmmlu-Korean-History-53",
            "kmmlu-Machine-Design-and-Manufacturing-449",
            "kmmlu-Math-205",
            "kmmlu-Political-Science-and-Sociology-198",
            "kmmlu-Psychology-892",
            "kmmlu-Public-Safety-152",
            "kmmlu-Railway-and-Automotive-Engineering-834"
          ]
        },
        "arc": {
          "n": 500,
          "accuracy": 0.514,
          "failures": 0,
          "accuracy_norm": 0.544,
          "delta_accuracy": {
            "delta": -0.004,
            "ci95": [
              -0.026,
              0.018
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "arc-Mercury_SC_414271",
            "arc-Mercury_7268328",
            "arc-Mercury_7211680",
            "arc-MCAS_2004_8_14",
            "arc-Mercury_7228725",
            "arc-Mercury_7007420",
            "arc-Mercury_7093048",
            "arc-ACTAAP_2010_7_12",
            "arc-Mercury_SC_407689",
            "arc-Mercury_7233765",
            "arc-MDSA_2007_8_42",
            "arc-Mercury_7212328",
            "arc-Mercury_7201740",
            "arc-Mercury_409295",
            "arc-Mercury_402364",
            "arc-Mercury_416536",
            "arc-Mercury_7215670"
          ],
          "improvements": [
            "arc-ACTAAP_2011_5_1",
            "arc-Mercury_7143308",
            "arc-Mercury_SC_400594",
            "arc-Mercury_416230",
            "arc-TIMSS_1995_8_K18",
            "arc-NYSEDREGENTS_2008_4_15",
            "arc-NYSEDREGENTS_2012_4_29",
            "arc-Mercury_417143",
            "arc-Mercury_SC_402164",
            "arc-MDSA_2009_5_20",
            "arc-Mercury_404132",
            "arc-Mercury_189070",
            "arc-NCEOGA_2013_8_28",
            "arc-AKDE&ED_2008_8_48",
            "arc-Mercury_SC_400708"
          ],
          "delta_accuracy_norm": {
            "delta": -0.018,
            "ci95": [
              -0.038,
              0.0
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          }
        },
        "ifeval": {
          "n": 541,
          "accuracy": 0.8096118299445472,
          "failures": 0,
          "instruction_strict": 0.8657074340527577,
          "truncated": 17,
          "delta_accuracy": {
            "delta": -0.0166358595194085,
            "ci95": [
              -0.04621072088724584,
              0.012939001848428836
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "ifeval-1075",
            "ifeval-1130",
            "ifeval-1379",
            "ifeval-1402",
            "ifeval-1436",
            "ifeval-1566",
            "ifeval-1670",
            "ifeval-1686",
            "ifeval-1713",
            "ifeval-179",
            "ifeval-1823",
            "ifeval-1825",
            "ifeval-1996",
            "ifeval-2084",
            "ifeval-2107",
            "ifeval-2250",
            "ifeval-2284",
            "ifeval-2311",
            "ifeval-279",
            "ifeval-2844",
            "ifeval-2912",
            "ifeval-2943",
            "ifeval-301",
            "ifeval-3048",
            "ifeval-3069",
            "ifeval-3071",
            "ifeval-3098",
            "ifeval-3109",
            "ifeval-3276",
            "ifeval-3326",
            "ifeval-3351",
            "ifeval-3362",
            "ifeval-3445",
            "ifeval-3453",
            "ifeval-3455",
            "ifeval-3479",
            "ifeval-3724"
          ],
          "improvements": [
            "ifeval-1021",
            "ifeval-1092",
            "ifeval-1122",
            "ifeval-1219",
            "ifeval-1265",
            "ifeval-127",
            "ifeval-1348",
            "ifeval-1372",
            "ifeval-1580",
            "ifeval-1629",
            "ifeval-1645",
            "ifeval-2266",
            "ifeval-2273",
            "ifeval-2299",
            "ifeval-2571",
            "ifeval-2591",
            "ifeval-2616",
            "ifeval-2617",
            "ifeval-2674",
            "ifeval-2736",
            "ifeval-2929",
            "ifeval-3035",
            "ifeval-3089",
            "ifeval-3114",
            "ifeval-331",
            "ifeval-3329",
            "ifeval-3409",
            "ifeval-3691"
          ]
        }
      },
      "speed": {
        "512": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.0330675,
            "min": 0.0327596,
            "max": 0.0332567
          },
          "prefill_seconds": {
            "median": 0.0330462,
            "min": 0.0327358,
            "max": 0.033232
          },
          "decode_seconds": {
            "median": 0.8875964,
            "min": 0.8829836,
            "max": 0.9592417
          },
          "seconds": {
            "median": 0.9208536,
            "min": 0.9160517,
            "max": 0.9923819
          },
          "prefill_tokens_per_second": {
            "median": 15493.460670213217,
            "min": 15406.836783822822,
            "max": 15640.36925934298
          },
          "decode_tokens_per_second": {
            "median": 287.2927380056972,
            "min": 265.8349819445923,
            "max": 288.79358574723244
          },
          "peak_device_vram_bytes": 7518580736
        },
        "4096": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.2629375,
            "min": 0.2624594,
            "max": 0.2783277
          },
          "prefill_seconds": {
            "median": 0.2629137,
            "min": 0.262434,
            "max": 0.2782977
          },
          "decode_seconds": {
            "median": 0.9103343,
            "min": 0.8922731,
            "max": 0.9743246
          },
          "seconds": {
            "median": 1.1761772,
            "min": 1.1551104,
            "max": 1.2405745
          },
          "prefill_tokens_per_second": {
            "median": 15579.256615383681,
            "min": 14718.051927845614,
            "max": 15607.733754010533
          },
          "decode_tokens_per_second": {
            "median": 280.11687574553656,
            "min": 261.7197595134106,
            "max": 285.7869412402996
          },
          "peak_device_vram_bytes": 7518580736
        }
      },
      "speed_source": "dedicated serial benchmark",
      "runtime_memory_breakdown": {
        "source": "llama.cpp CUDA0 buffer log",
        "model_mib": 3136.17,
        "kv_mib": 256.0,
        "recurrent_state_mib": 50.25,
        "compute_mib": 495.0
      },
      "block_diagnostics": {
        "scope": "six fixed validation documents, 256 tokens, cumulative error",
        "layer_mean_relative_l2": {
          "0": 0.066935607120169,
          "1": 0.08254003021591332,
          "2": 0.11221443734004773,
          "3": 0.1086405011262029,
          "4": 0.1191259817536946,
          "5": 0.12189637429590934,
          "6": 0.1239021199882645,
          "7": 0.13310617688155568,
          "8": 0.13458171267299124,
          "9": 0.1373012953631716,
          "10": 0.1355068317332434,
          "11": 0.1443239783354933,
          "12": 0.14671421864208473,
          "13": 0.1469534411375638,
          "14": 0.14865202202634462,
          "15": 0.15518710912462844,
          "16": 0.1583989896747106,
          "17": 0.15376421139806937,
          "18": 0.14641406686190905,
          "19": 0.15523886716075944,
          "20": 0.15056736005456164,
          "21": 0.14889112185060185,
          "22": 0.14890720506931343,
          "23": 0.1503209591401525,
          "24": 0.14933887836812265,
          "25": 0.1512918316579788,
          "26": 0.15762591483628324,
          "27": 0.16218375556570344,
          "28": 0.15676492841174577,
          "29": 0.16071543863577667,
          "30": 0.16840353121018434,
          "31": 0.16740250101368215
        }
      }
    },
    "native-KM": {
      "engine": "native",
      "candidate": "KM",
      "records": 4714,
      "failed_requests": [],
      "final_document_expected": 128,
      "final_document_present": 128,
      "final_document_failed": 0,
      "quality_complete": true,
      "nll": 1.8698531427248721,
      "ppl": 6.487343615720602,
      "scored_tokens": 130944,
      "domains": {
        "ko": {
          "documents": 64,
          "nll": 2.1160934507276865,
          "ppl": 8.298654977171498
        },
        "en": {
          "documents": 32,
          "nll": 2.179184125238734,
          "ppl": 8.839091724192476
        },
        "code": {
          "documents": 32,
          "nll": 1.068041544205382,
          "ppl": 2.9096754454942153
        }
      },
      "delta_nll": {
        "delta": 0.0183554999357027,
        "ci95": [
          0.016150725366032757,
          0.02042449977022933
        ],
        "resampling_unit": "paired document",
        "replicates": 2000
      },
      "ppl_ratio_vs_bf16": {
        "ratio": 1.0185249976078823,
        "ci95": [
          1.0162818533169937,
          1.0206345071943617
        ]
      },
      "document_delta_tail": {
        "mean": 0.018355499935702697,
        "median": 0.01692394045877729,
        "p95": 0.03945486485349074,
        "p99": 0.048711495942498746,
        "max": 0.061088601142256546
      },
      "worst20_documents": [
        {
          "id": "final-049",
          "domain": "code",
          "delta_nll": 0.061088601142256546,
          "nll": 0.5852325632764349
        },
        {
          "id": "final-012",
          "domain": "code",
          "delta_nll": 0.049407672854810375,
          "nll": 0.3242659135824352
        },
        {
          "id": "final-124",
          "domain": "ko",
          "delta_nll": 0.046829239846248746,
          "nll": 2.0941776115941666
        },
        {
          "id": "final-046",
          "domain": "code",
          "delta_nll": 0.04628819745892788,
          "nll": 1.32197961753998
        },
        {
          "id": "final-022",
          "domain": "ko",
          "delta_nll": 0.04345562407381191,
          "nll": 2.1537975690168363
        },
        {
          "id": "final-034",
          "domain": "ko",
          "delta_nll": 0.042037504139628326,
          "nll": 2.0228321322433813
        },
        {
          "id": "final-015",
          "domain": "ko",
          "delta_nll": 0.039921493092944904,
          "nll": 1.2557143029523974
        },
        {
          "id": "final-103",
          "domain": "en",
          "delta_nll": 0.038588269551647326,
          "nll": 2.1141448237771145
        },
        {
          "id": "final-009",
          "domain": "ko",
          "delta_nll": 0.03695002993118691,
          "nll": 2.2730098936404657
        },
        {
          "id": "final-126",
          "domain": "ko",
          "delta_nll": 0.03652271726803952,
          "nll": 2.4037907110900876
        },
        {
          "id": "final-063",
          "domain": "ko",
          "delta_nll": 0.03526436306020703,
          "nll": 1.9656501252710286
        },
        {
          "id": "final-033",
          "domain": "ko",
          "delta_nll": 0.034737506750652604,
          "nll": 2.3805538005415907
        },
        {
          "id": "final-062",
          "domain": "ko",
          "delta_nll": 0.03320915698318952,
          "nll": 2.1664839950567507
        },
        {
          "id": "final-010",
          "domain": "ko",
          "delta_nll": 0.03225646427612542,
          "nll": 2.4555286844605706
        },
        {
          "id": "final-003",
          "domain": "ko",
          "delta_nll": 0.03192312662065355,
          "nll": 1.4510748647628209
        },
        {
          "id": "final-089",
          "domain": "ko",
          "delta_nll": 0.03140233410873115,
          "nll": 2.7113327500963
        },
        {
          "id": "final-090",
          "domain": "code",
          "delta_nll": 0.031072726603585708,
          "nll": 0.41241853677999224
        },
        {
          "id": "final-083",
          "domain": "code",
          "delta_nll": 0.03068540944944374,
          "nll": 1.087048177123753
        },
        {
          "id": "final-002",
          "domain": "ko",
          "delta_nll": 0.030056604720066264,
          "nll": 1.9254702576781442
        },
        {
          "id": "final-081",
          "domain": "code",
          "delta_nll": 0.02982791185957918,
          "nll": 0.4702023801501512
        }
      ],
      "token_delta_tail": {
        "mean": 0.018355499935702454,
        "median": 0.00038291399268430837,
        "p95": 0.3769361662507298,
        "p99": 0.8090497631786076,
        "max": 6.117191639708771
      },
      "sampled_full_vocab_kl": {
        "mean": 0.024161768090731783,
        "median": 0.009382512824791842,
        "p95": 0.07177273292391109,
        "p99": 0.21476818007555457,
        "max": 3.3884346600221567
      },
      "kl_positions": 4096,
      "sampled_top1_agreement": 0.942138671875,
      "tasks": {
        "kmmlu": {
          "n": 500,
          "accuracy": 0.488,
          "failures": 0,
          "subject_scores": {
            "Accounting": {
              "n": 12,
              "accuracy": 0.5833333333333334
            },
            "Agricultural-Sciences": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Aviation-Engineering-and-Maintenance": {
              "n": 12,
              "accuracy": 0.5
            },
            "Biology": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Chemical-Engineering": {
              "n": 12,
              "accuracy": 0.5
            },
            "Chemistry": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Civil-Engineering": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Computer-Science": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Construction": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Criminal-Law": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Ecology": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Economics": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Education": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Electrical-Engineering": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Electronics-Engineering": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Energy-Management": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Environmental-Science": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Fashion": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Food-Processing": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Gas-Technology-and-Engineering": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Geomatics": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Health": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Industrial-Engineer": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Information-Technology": {
              "n": 11,
              "accuracy": 1.0
            },
            "Interior-Architecture-and-Design": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Korean-History": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Law": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Machine-Design-and-Manufacturing": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Management": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Maritime-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Marketing": {
              "n": 11,
              "accuracy": 0.9090909090909091
            },
            "Materials-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Math": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Mechanical-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Nondestructive-Testing": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Patent": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Political-Science-and-Sociology": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Psychology": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Public-Safety": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Railway-and-Automotive-Engineering": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Real-Estate": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Refrigerating-Machinery": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Social-Welfare": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Taxation": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Telecommunications-and-Wireless-Technology": {
              "n": 11,
              "accuracy": 0.6363636363636364
            }
          },
          "macro_accuracy": 0.48804713804713795,
          "delta_accuracy": {
            "delta": 0.006,
            "ci95": [
              -0.018,
              0.028
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "kmmlu-Agricultural-Sciences-226",
            "kmmlu-Chemistry-486",
            "kmmlu-Civil-Engineering-837",
            "kmmlu-Economics-129",
            "kmmlu-Electrical-Engineering-481",
            "kmmlu-Electronics-Engineering-83",
            "kmmlu-Energy-Management-645",
            "kmmlu-Energy-Management-431",
            "kmmlu-Geomatics-629",
            "kmmlu-Industrial-Engineer-130",
            "kmmlu-Interior-Architecture-and-Design-849",
            "kmmlu-Interior-Architecture-and-Design-560",
            "kmmlu-Machine-Design-and-Manufacturing-699",
            "kmmlu-Nondestructive-Testing-286",
            "kmmlu-Patent-99",
            "kmmlu-Political-Science-and-Sociology-76",
            "kmmlu-Refrigerating-Machinery-797"
          ],
          "improvements": [
            "kmmlu-Biology-731",
            "kmmlu-Chemical-Engineering-673",
            "kmmlu-Chemistry-130",
            "kmmlu-Criminal-Law-172",
            "kmmlu-Economics-26",
            "kmmlu-Electronics-Engineering-202",
            "kmmlu-Geomatics-15",
            "kmmlu-Industrial-Engineer-141",
            "kmmlu-Interior-Architecture-and-Design-723",
            "kmmlu-Machine-Design-and-Manufacturing-449",
            "kmmlu-Math-205",
            "kmmlu-Patent-48",
            "kmmlu-Political-Science-and-Sociology-198",
            "kmmlu-Psychology-892",
            "kmmlu-Psychology-939",
            "kmmlu-Public-Safety-152",
            "kmmlu-Railway-and-Automotive-Engineering-834",
            "kmmlu-Real-Estate-184",
            "kmmlu-Refrigerating-Machinery-302",
            "kmmlu-Social-Welfare-19"
          ]
        },
        "arc": {
          "n": 500,
          "accuracy": 0.514,
          "failures": 0,
          "accuracy_norm": 0.538,
          "delta_accuracy": {
            "delta": -0.004,
            "ci95": [
              -0.026,
              0.018
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "arc-Mercury_7025060",
            "arc-Mercury_SC_414271",
            "arc-Mercury_7268328",
            "arc-Mercury_7211680",
            "arc-MCAS_2004_8_14",
            "arc-Mercury_7007420",
            "arc-Mercury_7093048",
            "arc-ACTAAP_2010_7_12",
            "arc-Mercury_SC_407689",
            "arc-Mercury_187198",
            "arc-NCEOGA_2013_8_42",
            "arc-Mercury_7201740",
            "arc-Mercury_409295",
            "arc-Mercury_402364",
            "arc-Mercury_416536",
            "arc-Mercury_7215670"
          ],
          "improvements": [
            "arc-Mercury_7143308",
            "arc-Mercury_SC_400594",
            "arc-TIMSS_1995_8_K18",
            "arc-NYSEDREGENTS_2008_4_15",
            "arc-NYSEDREGENTS_2012_4_29",
            "arc-Mercury_SC_409241",
            "arc-Mercury_7136885",
            "arc-Mercury_417143",
            "arc-Mercury_SC_402164",
            "arc-MDSA_2009_5_20",
            "arc-Mercury_189070",
            "arc-NCEOGA_2013_8_28",
            "arc-Mercury_SC_400708",
            "arc-Mercury_7056298"
          ],
          "delta_accuracy_norm": {
            "delta": -0.024,
            "ci95": [
              -0.042,
              -0.006
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          }
        },
        "ifeval": {
          "n": 541,
          "accuracy": 0.7985212569316081,
          "failures": 0,
          "instruction_strict": 0.8585131894484412,
          "truncated": 21,
          "delta_accuracy": {
            "delta": -0.027726432532347505,
            "ci95": [
              -0.05545286506469501,
              0.0
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "ifeval-1087",
            "ifeval-1129",
            "ifeval-1130",
            "ifeval-1379",
            "ifeval-1566",
            "ifeval-1609",
            "ifeval-1670",
            "ifeval-1686",
            "ifeval-179",
            "ifeval-1823",
            "ifeval-1837",
            "ifeval-2034",
            "ifeval-2063",
            "ifeval-2139",
            "ifeval-2180",
            "ifeval-2275",
            "ifeval-2303",
            "ifeval-2311",
            "ifeval-2416",
            "ifeval-2439",
            "ifeval-2605",
            "ifeval-279",
            "ifeval-2844",
            "ifeval-2859",
            "ifeval-2871",
            "ifeval-2918",
            "ifeval-2943",
            "ifeval-3098",
            "ifeval-3109",
            "ifeval-3223",
            "ifeval-3276",
            "ifeval-3287",
            "ifeval-3351",
            "ifeval-3362",
            "ifeval-3429",
            "ifeval-3453",
            "ifeval-3455",
            "ifeval-3479",
            "ifeval-3724"
          ],
          "improvements": [
            "ifeval-1021",
            "ifeval-1092",
            "ifeval-1122",
            "ifeval-1219",
            "ifeval-1265",
            "ifeval-127",
            "ifeval-1348",
            "ifeval-1372",
            "ifeval-1580",
            "ifeval-1582",
            "ifeval-1629",
            "ifeval-1645",
            "ifeval-1944",
            "ifeval-2266",
            "ifeval-2299",
            "ifeval-2337",
            "ifeval-2591",
            "ifeval-2616",
            "ifeval-2617",
            "ifeval-2736",
            "ifeval-2929",
            "ifeval-3114",
            "ifeval-3409",
            "ifeval-3456"
          ]
        }
      },
      "speed": {
        "512": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.0339146,
            "min": 0.03385,
            "max": 0.0419931
          },
          "prefill_seconds": {
            "median": 0.0338926,
            "min": 0.0338276,
            "max": 0.0419684
          },
          "decode_seconds": {
            "median": 0.9326449,
            "min": 0.8984739,
            "max": 0.9876433
          },
          "seconds": {
            "median": 0.974639,
            "min": 0.9323247,
            "max": 1.0214961
          },
          "prefill_tokens_per_second": {
            "median": 15106.542431091151,
            "min": 12199.654978507639,
            "max": 15135.569771429247
          },
          "decode_tokens_per_second": {
            "median": 273.41595927882094,
            "min": 258.1903810819149,
            "max": 283.8145882701768
          },
          "peak_device_vram_bytes": 7735070720
        },
        "4096": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.270575,
            "min": 0.2692284,
            "max": 0.3358204
          },
          "prefill_seconds": {
            "median": 0.2705501,
            "min": 0.2692013,
            "max": 0.3357967
          },
          "decode_seconds": {
            "median": 0.9245138,
            "min": 0.9161905,
            "max": 1.0056731
          },
          "seconds": {
            "median": 1.1996949,
            "min": 1.1883828,
            "max": 1.2755291
          },
          "prefill_tokens_per_second": {
            "median": 15139.524990011092,
            "min": 12197.85661979406,
            "max": 15215.379717705673
          },
          "decode_tokens_per_second": {
            "median": 275.8206529745689,
            "min": 253.56152014009322,
            "max": 278.3263960933889
          },
          "peak_device_vram_bytes": 7735070720
        }
      },
      "speed_source": "dedicated serial benchmark",
      "runtime_memory_breakdown": {
        "source": "llama.cpp CUDA0 buffer log",
        "model_mib": 3293.43,
        "kv_mib": 256.0,
        "recurrent_state_mib": 50.25,
        "compute_mib": 495.0
      },
      "block_diagnostics": {
        "scope": "six fixed validation documents, 256 tokens, cumulative error",
        "layer_mean_relative_l2": {
          "0": 0.06221998385965822,
          "1": 0.0736281802169728,
          "2": 0.09711888425739844,
          "3": 0.09315695809951617,
          "4": 0.10560414343263456,
          "5": 0.11069646739596915,
          "6": 0.10908242706046767,
          "7": 0.1205631162176491,
          "8": 0.1233846479318429,
          "9": 0.12405128962390997,
          "10": 0.12408944420636558,
          "11": 0.1338146160702853,
          "12": 0.1343382014730076,
          "13": 0.13585275683917755,
          "14": 0.13853996215654016,
          "15": 0.14370073753881193,
          "16": 0.14785750032340572,
          "17": 0.14457445721315573,
          "18": 0.13532864274733278,
          "19": 0.14613327878598575,
          "20": 0.14250442456763796,
          "21": 0.13964477766013275,
          "22": 0.1437853252215492,
          "23": 0.14390431999820316,
          "24": 0.14250248377598732,
          "25": 0.14520986439265834,
          "26": 0.15554735659349533,
          "27": 0.1550027367062706,
          "28": 0.1483145894470747,
          "29": 0.15122062429009922,
          "30": 0.15901377426428184,
          "31": 0.15746219683377075
        }
      }
    },
    "vllm-A0": {
      "engine": "vllm",
      "candidate": "A0",
      "records": 4714,
      "failed_requests": [],
      "final_document_expected": 128,
      "final_document_present": 128,
      "final_document_failed": 0,
      "quality_complete": true,
      "nll": 1.933766899369248,
      "ppl": 6.915511274256328,
      "scored_tokens": 130944,
      "domains": {
        "ko": {
          "documents": 64,
          "nll": 2.1874108605731073,
          "ppl": 8.91210852554423
        },
        "en": {
          "documents": 32,
          "nll": 2.2382833058531633,
          "ppl": 9.377219643941181
        },
        "code": {
          "documents": 32,
          "nll": 1.1219625704776142,
          "ppl": 3.070875101957011
        }
      },
      "delta_nll": {
        "delta": 0.08193356100092503,
        "ci95": [
          0.07663695831162505,
          0.08749693774992422
        ],
        "resampling_unit": "paired document",
        "replicates": 2000
      },
      "ppl_ratio_vs_bf16": {
        "ratio": 1.0853836956276282,
        "ci95": [
          1.0796500472594357,
          1.091438922178912
        ]
      },
      "document_delta_tail": {
        "mean": 0.08193356100092503,
        "median": 0.08127681749337565,
        "p95": 0.13445226782526293,
        "p99": 0.17382731818548527,
        "max": 0.18086348787961604
      },
      "worst20_documents": [
        {
          "id": "final-109",
          "domain": "en",
          "delta_nll": 0.18086348787961604,
          "nll": 2.278217233155557
        },
        {
          "id": "final-049",
          "domain": "code",
          "delta_nll": 0.17546424062354768,
          "nll": 0.7017778313493083
        },
        {
          "id": "final-035",
          "domain": "ko",
          "delta_nll": 0.16940156492702016,
          "nll": 2.2589829856584447
        },
        {
          "id": "final-103",
          "domain": "en",
          "delta_nll": 0.1492712258299406,
          "nll": 2.224580187152183
        },
        {
          "id": "final-009",
          "domain": "ko",
          "delta_nll": 0.14808116910993888,
          "nll": 2.38311079912873
        },
        {
          "id": "final-075",
          "domain": "ko",
          "delta_nll": 0.1374624848805378,
          "nll": 2.072410943751995
        },
        {
          "id": "final-018",
          "domain": "ko",
          "delta_nll": 0.1348537893556687,
          "nll": 2.4294281420680353
        },
        {
          "id": "final-003",
          "domain": "ko",
          "delta_nll": 0.13370658498308086,
          "nll": 1.5522493284111323
        },
        {
          "id": "final-063",
          "domain": "ko",
          "delta_nll": 0.13283096110976852,
          "nll": 2.0636811395313632
        },
        {
          "id": "final-012",
          "domain": "code",
          "delta_nll": 0.12851868521360962,
          "nll": 0.40393711445697955
        },
        {
          "id": "final-001",
          "domain": "code",
          "delta_nll": 0.12633690652101745,
          "nll": 0.3265917188627491
        },
        {
          "id": "final-039",
          "domain": "ko",
          "delta_nll": 0.12100410085645907,
          "nll": 2.364925483530263
        },
        {
          "id": "final-036",
          "domain": "ko",
          "delta_nll": 0.11785046078056192,
          "nll": 2.3910732522763407
        },
        {
          "id": "final-113",
          "domain": "ko",
          "delta_nll": 0.11768524093174021,
          "nll": 2.367358109734549
        },
        {
          "id": "final-032",
          "domain": "ko",
          "delta_nll": 0.1170312009950744,
          "nll": 1.9540231333082116
        },
        {
          "id": "final-033",
          "domain": "ko",
          "delta_nll": 0.11413480868680381,
          "nll": 2.4609519868459198
        },
        {
          "id": "final-046",
          "domain": "code",
          "delta_nll": 0.1134328493972261,
          "nll": 1.3898805659945797
        },
        {
          "id": "final-047",
          "domain": "ko",
          "delta_nll": 0.11254767220010566,
          "nll": 1.6434353927459193
        },
        {
          "id": "final-019",
          "domain": "en",
          "delta_nll": 0.11151946451968953,
          "nll": 2.011857288835099
        },
        {
          "id": "final-107",
          "domain": "ko",
          "delta_nll": 0.10808832735124835,
          "nll": 2.130467384430413
        }
      ],
      "token_delta_tail": {
        "mean": 0.08193356100092504,
        "median": 0.0050355298444628716,
        "p95": 0.8673044264316554,
        "p99": 1.8089004468917764,
        "max": 15.382938370108604
      },
      "sampled_full_vocab_kl": null,
      "kl_positions": 0,
      "tasks": {
        "kmmlu": {
          "n": 500,
          "accuracy": 0.438,
          "failures": 0,
          "subject_scores": {
            "Accounting": {
              "n": 12,
              "accuracy": 0.5
            },
            "Agricultural-Sciences": {
              "n": 12,
              "accuracy": 0.16666666666666666
            },
            "Aviation-Engineering-and-Maintenance": {
              "n": 12,
              "accuracy": 0.5833333333333334
            },
            "Biology": {
              "n": 12,
              "accuracy": 0.3333333333333333
            },
            "Chemical-Engineering": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Chemistry": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Civil-Engineering": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Computer-Science": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Construction": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Criminal-Law": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Ecology": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Economics": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Education": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Electrical-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Electronics-Engineering": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Energy-Management": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Environmental-Science": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Fashion": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Food-Processing": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Gas-Technology-and-Engineering": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Geomatics": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Health": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Industrial-Engineer": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Information-Technology": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Interior-Architecture-and-Design": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Korean-History": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Law": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Machine-Design-and-Manufacturing": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Management": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Maritime-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Marketing": {
              "n": 11,
              "accuracy": 1.0
            },
            "Materials-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Math": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Mechanical-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Nondestructive-Testing": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Patent": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Political-Science-and-Sociology": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Psychology": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Public-Safety": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Railway-and-Automotive-Engineering": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Real-Estate": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Refrigerating-Machinery": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Social-Welfare": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Taxation": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Telecommunications-and-Wireless-Technology": {
              "n": 11,
              "accuracy": 0.5454545454545454
            }
          },
          "macro_accuracy": 0.4383838383838384,
          "delta_accuracy": {
            "delta": -0.05,
            "ci95": [
              -0.09,
              -0.014
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "kmmlu-Accounting-96",
            "kmmlu-Agricultural-Sciences-999",
            "kmmlu-Agricultural-Sciences-736",
            "kmmlu-Agricultural-Sciences-660",
            "kmmlu-Agricultural-Sciences-226",
            "kmmlu-Aviation-Engineering-and-Maintenance-899",
            "kmmlu-Biology-731",
            "kmmlu-Chemical-Engineering-403",
            "kmmlu-Chemistry-367",
            "kmmlu-Criminal-Law-73",
            "kmmlu-Ecology-998",
            "kmmlu-Ecology-577",
            "kmmlu-Economics-38",
            "kmmlu-Economics-129",
            "kmmlu-Economics-106",
            "kmmlu-Education-20",
            "kmmlu-Electrical-Engineering-890",
            "kmmlu-Electrical-Engineering-135",
            "kmmlu-Electronics-Engineering-294",
            "kmmlu-Electronics-Engineering-810",
            "kmmlu-Electronics-Engineering-83",
            "kmmlu-Electronics-Engineering-78",
            "kmmlu-Energy-Management-155",
            "kmmlu-Environmental-Science-192",
            "kmmlu-Environmental-Science-342",
            "kmmlu-Fashion-419",
            "kmmlu-Fashion-249",
            "kmmlu-Food-Processing-225",
            "kmmlu-Gas-Technology-and-Engineering-361",
            "kmmlu-Geomatics-272",
            "kmmlu-Information-Technology-532",
            "kmmlu-Information-Technology-635",
            "kmmlu-Interior-Architecture-and-Design-849",
            "kmmlu-Interior-Architecture-and-Design-560",
            "kmmlu-Interior-Architecture-and-Design-926",
            "kmmlu-Korean-History-93",
            "kmmlu-Korean-History-28",
            "kmmlu-Law-755",
            "kmmlu-Machine-Design-and-Manufacturing-690",
            "kmmlu-Machine-Design-and-Manufacturing-815",
            "kmmlu-Machine-Design-and-Manufacturing-345",
            "kmmlu-Machine-Design-and-Manufacturing-326",
            "kmmlu-Management-192",
            "kmmlu-Maritime-Engineering-145",
            "kmmlu-Maritime-Engineering-115",
            "kmmlu-Materials-Engineering-35",
            "kmmlu-Math-208",
            "kmmlu-Math-246",
            "kmmlu-Nondestructive-Testing-500",
            "kmmlu-Nondestructive-Testing-820",
            "kmmlu-Patent-99",
            "kmmlu-Public-Safety-941",
            "kmmlu-Public-Safety-262",
            "kmmlu-Railway-and-Automotive-Engineering-685",
            "kmmlu-Refrigerating-Machinery-797",
            "kmmlu-Refrigerating-Machinery-911",
            "kmmlu-Social-Welfare-952",
            "kmmlu-Social-Welfare-383",
            "kmmlu-Telecommunications-and-Wireless-Technology-118",
            "kmmlu-Telecommunications-and-Wireless-Technology-456"
          ],
          "improvements": [
            "kmmlu-Aviation-Engineering-and-Maintenance-285",
            "kmmlu-Aviation-Engineering-and-Maintenance-380",
            "kmmlu-Chemical-Engineering-837",
            "kmmlu-Civil-Engineering-594",
            "kmmlu-Civil-Engineering-924",
            "kmmlu-Construction-396",
            "kmmlu-Construction-787",
            "kmmlu-Criminal-Law-40",
            "kmmlu-Economics-74",
            "kmmlu-Electrical-Engineering-118",
            "kmmlu-Electrical-Engineering-386",
            "kmmlu-Environmental-Science-835",
            "kmmlu-Food-Processing-919",
            "kmmlu-Gas-Technology-and-Engineering-705",
            "kmmlu-Geomatics-951",
            "kmmlu-Industrial-Engineer-927",
            "kmmlu-Interior-Architecture-and-Design-723",
            "kmmlu-Interior-Architecture-and-Design-947",
            "kmmlu-Korean-History-89",
            "kmmlu-Law-277",
            "kmmlu-Machine-Design-and-Manufacturing-449",
            "kmmlu-Maritime-Engineering-24",
            "kmmlu-Maritime-Engineering-234",
            "kmmlu-Marketing-80",
            "kmmlu-Materials-Engineering-321",
            "kmmlu-Math-144",
            "kmmlu-Patent-48",
            "kmmlu-Political-Science-and-Sociology-213",
            "kmmlu-Public-Safety-152",
            "kmmlu-Railway-and-Automotive-Engineering-177",
            "kmmlu-Real-Estate-172",
            "kmmlu-Social-Welfare-706",
            "kmmlu-Taxation-71",
            "kmmlu-Taxation-148",
            "kmmlu-Telecommunications-and-Wireless-Technology-713"
          ]
        },
        "arc": {
          "n": 500,
          "accuracy": 0.494,
          "failures": 0,
          "accuracy_norm": 0.532,
          "delta_accuracy": {
            "delta": -0.028,
            "ci95": [
              -0.056,
              -0.002
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "arc-Mercury_7025060",
            "arc-Mercury_7162400",
            "arc-Mercury_SC_414271",
            "arc-Mercury_7086608",
            "arc-MCAS_2005_9_6",
            "arc-Mercury_7007420",
            "arc-Mercury_7093048",
            "arc-Mercury_SC_407689",
            "arc-Mercury_SC_406451",
            "arc-Mercury_SC_415453",
            "arc-TIMSS_2007_8_pg113",
            "arc-Mercury_417138",
            "arc-Mercury_SC_409241",
            "arc-Mercury_SC_400839",
            "arc-Mercury_7136885",
            "arc-MCAS_2004_9_2",
            "arc-Mercury_7221043",
            "arc-Mercury_7042945",
            "arc-Mercury_7227710",
            "arc-Mercury_7012740",
            "arc-MDSA_2007_8_42",
            "arc-Mercury_7270113",
            "arc-Mercury_SC_406720",
            "arc-NCEOGA_2013_8_42",
            "arc-MCAS_2010_8_12005",
            "arc-Mercury_7212328",
            "arc-Mercury_7201740",
            "arc-Mercury_409295",
            "arc-Mercury_7171430",
            "arc-MEAP_2005_5_12"
          ],
          "improvements": [
            "arc-ACTAAP_2011_5_1",
            "arc-MEAP_2005_8_39",
            "arc-Mercury_SC_407613",
            "arc-Mercury_7206395",
            "arc-MCAS_2013_5_29411",
            "arc-Mercury_7210210",
            "arc-MDSA_2008_8_27",
            "arc-Mercury_SC_LBS10337",
            "arc-Mercury_7216248",
            "arc-Mercury_417143",
            "arc-Mercury_417118",
            "arc-Mercury_7268240",
            "arc-Mercury_7283833",
            "arc-Mercury_7185343",
            "arc-NCEOGA_2013_8_28",
            "arc-Mercury_SC_400708"
          ],
          "delta_accuracy_norm": {
            "delta": -0.026,
            "ci95": [
              -0.05,
              -0.002
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          }
        },
        "ifeval": {
          "n": 541,
          "accuracy": 0.7818853974121996,
          "failures": 0,
          "instruction_strict": 0.8465227817745803,
          "truncated": 49,
          "delta_accuracy": {
            "delta": -0.04990757855822551,
            "ci95": [
              -0.07948243992606285,
              -0.02033271719038817
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "ifeval-1000",
            "ifeval-1122",
            "ifeval-1129",
            "ifeval-1216",
            "ifeval-1219",
            "ifeval-122",
            "ifeval-1233",
            "ifeval-1242",
            "ifeval-1332",
            "ifeval-1342",
            "ifeval-1348",
            "ifeval-1392",
            "ifeval-1481",
            "ifeval-1670",
            "ifeval-1691",
            "ifeval-1713",
            "ifeval-179",
            "ifeval-1825",
            "ifeval-1893",
            "ifeval-1908",
            "ifeval-1928",
            "ifeval-1967",
            "ifeval-2035",
            "ifeval-2063",
            "ifeval-2118",
            "ifeval-2250",
            "ifeval-2386",
            "ifeval-2416",
            "ifeval-2439",
            "ifeval-2563",
            "ifeval-2591",
            "ifeval-2596",
            "ifeval-2605",
            "ifeval-2716",
            "ifeval-279",
            "ifeval-2811",
            "ifeval-2820",
            "ifeval-2905",
            "ifeval-2908",
            "ifeval-2912",
            "ifeval-2918",
            "ifeval-2929",
            "ifeval-3069",
            "ifeval-3098",
            "ifeval-3186",
            "ifeval-3223",
            "ifeval-3276",
            "ifeval-3324",
            "ifeval-3442",
            "ifeval-3445",
            "ifeval-3455",
            "ifeval-3502"
          ],
          "improvements": [
            "ifeval-1092",
            "ifeval-1203",
            "ifeval-1389",
            "ifeval-1629",
            "ifeval-2034",
            "ifeval-2078",
            "ifeval-2084",
            "ifeval-2180",
            "ifeval-2266",
            "ifeval-2299",
            "ifeval-2350",
            "ifeval-2617",
            "ifeval-2662",
            "ifeval-2736",
            "ifeval-2941",
            "ifeval-3089",
            "ifeval-3109",
            "ifeval-3256",
            "ifeval-3329",
            "ifeval-3409",
            "ifeval-3428",
            "ifeval-3429",
            "ifeval-3453",
            "ifeval-3484",
            "ifeval-3691"
          ]
        }
      },
      "speed": {
        "512": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.045508384704589844,
            "min": 0.044849395751953125,
            "max": 0.09570884704589844
          },
          "prefill_seconds": {
            "median": 0.04440697999999088,
            "min": 0.04364459399999987,
            "max": 0.09454871199999104
          },
          "decode_seconds": {
            "median": 5.094352267999994,
            "min": 4.720216910999994,
            "max": 7.162745676
          },
          "seconds": {
            "median": 5.1390060800000015,
            "min": 4.765158537000005,
            "max": 7.2196227419999985
          },
          "prefill_tokens_per_second": {
            "median": 11529.718976613702,
            "min": 5415.198040984932,
            "max": 11731.120697330844
          },
          "decode_tokens_per_second": {
            "median": 50.055431306110115,
            "min": 35.60087311970617,
            "max": 54.02294106564212
          },
          "peak_device_vram_bytes": 30523822080
        },
        "4096": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.2110755443572998,
            "min": 0.20933222770690918,
            "max": 0.2458353042602539
          },
          "prefill_seconds": {
            "median": 0.20952049599999611,
            "min": 0.20781107700000234,
            "max": 0.24418523199999242
          },
          "decode_seconds": {
            "median": 5.667101024000004,
            "min": 4.674583480999999,
            "max": 6.1266377329999955
          },
          "seconds": {
            "median": 5.877772003000004,
            "min": 4.883624401000006,
            "max": 6.372081207999997
          },
          "prefill_tokens_per_second": {
            "median": 19549.40007396735,
            "min": 16774.151190273977,
            "max": 19710.210154004224
          },
          "decode_tokens_per_second": {
            "median": 44.99655095613835,
            "min": 41.621524090854905,
            "max": 54.55031470428457
          },
          "peak_device_vram_bytes": 31200645120
        }
      },
      "speed_source": "dedicated serial benchmark",
      "runtime_memory_breakdown": {
        "source": "vLLM startup profile, rounded GiB",
        "weight_gib": 3.11,
        "profile_peak_activation_gib": 0.42,
        "non_torch_gib": 0.15,
        "reserved_kv_cache_gib": 21.77
      }
    },
    "vllm-B0": {
      "engine": "vllm",
      "candidate": "B0",
      "records": 4714,
      "failed_requests": [],
      "final_document_expected": 128,
      "final_document_present": 128,
      "final_document_failed": 0,
      "quality_complete": true,
      "nll": 1.851833338368323,
      "ppl": 6.371489918371587,
      "scored_tokens": 130944,
      "domains": {
        "ko": {
          "documents": 64,
          "nll": 2.0939456000989347,
          "ppl": 8.116878021237127
        },
        "en": {
          "documents": 32,
          "nll": 2.1695132429391744,
          "ppl": 8.754021921535749
        },
        "code": {
          "documents": 32,
          "nll": 1.0499289103362484,
          "ppl": 2.8574479758268034
        }
      },
      "delta_nll": {
        "delta": 0.0,
        "ci95": [
          0.0,
          0.0
        ],
        "resampling_unit": "paired document",
        "replicates": 2000
      },
      "ppl_ratio_vs_bf16": {
        "ratio": 1.0,
        "ci95": [
          1.0,
          1.0
        ]
      },
      "document_delta_tail": {
        "mean": 0.0,
        "median": 0.0,
        "p95": 0.0,
        "p99": 0.0,
        "max": 0.0
      },
      "worst20_documents": [
        {
          "id": "final-000",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 2.2103152485675994
        },
        {
          "id": "final-001",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 0.20025481234173167
        },
        {
          "id": "final-002",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.8941187276930938
        },
        {
          "id": "final-003",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.4185427434280515
        },
        {
          "id": "final-004",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 1.177325957679266
        },
        {
          "id": "final-005",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 2.2068631005641315
        },
        {
          "id": "final-006",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 0.7944584573940342
        },
        {
          "id": "final-007",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.93964734846893
        },
        {
          "id": "final-008",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 2.29841878664696
        },
        {
          "id": "final-009",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.2350296300187913
        },
        {
          "id": "final-010",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.4237570619428
        },
        {
          "id": "final-011",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.5659503719011485
        },
        {
          "id": "final-012",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 0.2754184292433699
        },
        {
          "id": "final-013",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 2.4716811972016886
        },
        {
          "id": "final-014",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.636997431389201
        },
        {
          "id": "final-015",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 1.2153571033151187
        },
        {
          "id": "final-016",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.3393618316037705
        },
        {
          "id": "final-017",
          "domain": "code",
          "delta_nll": 0.0,
          "nll": 1.0857602734252005
        },
        {
          "id": "final-018",
          "domain": "ko",
          "delta_nll": 0.0,
          "nll": 2.2945743527123668
        },
        {
          "id": "final-019",
          "domain": "en",
          "delta_nll": 0.0,
          "nll": 1.9003378243154094
        }
      ],
      "token_delta_tail": {
        "mean": 0.0,
        "median": 0.0,
        "p95": 0.0,
        "p99": 0.0,
        "max": 0.0
      },
      "sampled_full_vocab_kl": null,
      "kl_positions": 0,
      "tasks": {
        "kmmlu": {
          "n": 500,
          "accuracy": 0.488,
          "failures": 0,
          "subject_scores": {
            "Accounting": {
              "n": 12,
              "accuracy": 0.5833333333333334
            },
            "Agricultural-Sciences": {
              "n": 12,
              "accuracy": 0.5
            },
            "Aviation-Engineering-and-Maintenance": {
              "n": 12,
              "accuracy": 0.5
            },
            "Biology": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Chemical-Engineering": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Chemistry": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Civil-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Computer-Science": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Construction": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Criminal-Law": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Ecology": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Economics": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Education": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Electrical-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Electronics-Engineering": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Energy-Management": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Environmental-Science": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Fashion": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Food-Processing": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Gas-Technology-and-Engineering": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Geomatics": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Health": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Industrial-Engineer": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Information-Technology": {
              "n": 11,
              "accuracy": 1.0
            },
            "Interior-Architecture-and-Design": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Korean-History": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Law": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Machine-Design-and-Manufacturing": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Management": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Maritime-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Marketing": {
              "n": 11,
              "accuracy": 0.9090909090909091
            },
            "Materials-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Math": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Mechanical-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Nondestructive-Testing": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Patent": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Political-Science-and-Sociology": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Psychology": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Public-Safety": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Railway-and-Automotive-Engineering": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Real-Estate": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Refrigerating-Machinery": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Social-Welfare": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Taxation": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Telecommunications-and-Wireless-Technology": {
              "n": 11,
              "accuracy": 0.6363636363636364
            }
          },
          "macro_accuracy": 0.48804713804713795,
          "delta_accuracy": {
            "delta": 0.0,
            "ci95": [
              0.0,
              0.0
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [],
          "improvements": []
        },
        "arc": {
          "n": 500,
          "accuracy": 0.522,
          "failures": 0,
          "accuracy_norm": 0.558,
          "delta_accuracy": {
            "delta": 0.0,
            "ci95": [
              0.0,
              0.0
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [],
          "improvements": [],
          "delta_accuracy_norm": {
            "delta": 0.0,
            "ci95": [
              0.0,
              0.0
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          }
        },
        "ifeval": {
          "n": 541,
          "accuracy": 0.8317929759704251,
          "failures": 0,
          "instruction_strict": 0.8836930455635491,
          "truncated": 14,
          "delta_accuracy": {
            "delta": 0.0,
            "ci95": [
              0.0,
              0.0
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [],
          "improvements": []
        }
      },
      "speed": {
        "512": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.044440269470214844,
            "min": 0.043544769287109375,
            "max": 0.046628475189208984
          },
          "prefill_seconds": {
            "median": 0.04305331400001933,
            "min": 0.04249180099998284,
            "max": 0.04555048100002068
          },
          "decode_seconds": {
            "median": 4.408005844999991,
            "min": 3.7740833859999725,
            "max": 5.711164961999998
          },
          "seconds": {
            "median": 4.4517488000000185,
            "min": 3.8205934830000388,
            "max": 5.75528235500002
          },
          "prefill_tokens_per_second": {
            "median": 11892.232035837476,
            "min": 11240.276474792166,
            "max": 12049.383362221026
          },
          "decode_tokens_per_second": {
            "median": 57.849288083237674,
            "min": 44.64938444199681,
            "max": 67.56607470463607
          },
          "peak_device_vram_bytes": 30463041536
        },
        "4096": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.20170092582702637,
            "min": 0.1910536289215088,
            "max": 0.21390032768249512
          },
          "prefill_seconds": {
            "median": 0.2003497110000012,
            "min": 0.189843097999983,
            "max": 0.21237152999998443
          },
          "decode_seconds": {
            "median": 4.531537686999968,
            "min": 4.341469549999999,
            "max": 5.642180597999982
          },
          "seconds": {
            "median": 4.725557533000028,
            "min": 4.554985737999971,
            "max": 5.833118721000005
          },
          "prefill_tokens_per_second": {
            "median": 20444.25210076782,
            "min": 19286.954329520064,
            "max": 21575.711959780423
          },
          "decode_tokens_per_second": {
            "median": 56.2722893669276,
            "min": 45.19529206321248,
            "max": 58.735872050513414
          },
          "peak_device_vram_bytes": 31085576192
        }
      },
      "speed_source": "dedicated serial benchmark",
      "runtime_memory_breakdown": {
        "source": "vLLM startup profile, rounded GiB",
        "weight_gib": 7.99,
        "profile_peak_activation_gib": 0.41,
        "non_torch_gib": 0.15,
        "reserved_kv_cache_gib": 16.9
      }
    },
    "vllm-G0": {
      "engine": "vllm",
      "candidate": "G0",
      "records": 4714,
      "failed_requests": [],
      "final_document_expected": 128,
      "final_document_present": 128,
      "final_document_failed": 0,
      "quality_complete": true,
      "nll": 1.8883687908364546,
      "ppl": 6.6085799079587035,
      "scored_tokens": 130944,
      "domains": {
        "ko": {
          "documents": 64,
          "nll": 2.1323308151056413,
          "ppl": 8.43450318560508
        },
        "en": {
          "documents": 32,
          "nll": 2.2066754175675585,
          "ppl": 9.085460763994856
        },
        "code": {
          "documents": 32,
          "nll": 1.0821381155669765,
          "ppl": 2.950982351910591
        }
      },
      "delta_nll": {
        "delta": 0.0365354524681315,
        "ci95": [
          0.03341807698834355,
          0.03985762477482175
        ],
        "resampling_unit": "paired document",
        "replicates": 2000
      },
      "ppl_ratio_vs_bf16": {
        "ratio": 1.037211075058518,
        "ci95": [
          1.0339827332769476,
          1.040662599072528
        ]
      },
      "document_delta_tail": {
        "mean": 0.03653545246813151,
        "median": 0.035811706725006574,
        "p95": 0.06272625737042016,
        "p99": 0.09966297742808866,
        "max": 0.13673098418300175
      },
      "worst20_documents": [
        {
          "id": "final-049",
          "domain": "code",
          "delta_nll": 0.13673098418300175,
          "nll": 0.6630445749087623
        },
        {
          "id": "final-103",
          "domain": "en",
          "delta_nll": 0.10774578183689297,
          "nll": 2.1830547431591354
        },
        {
          "id": "final-035",
          "domain": "ko",
          "delta_nll": 0.07780946921169136,
          "nll": 2.167390889943116
        },
        {
          "id": "final-036",
          "domain": "ko",
          "delta_nll": 0.07711811391275634,
          "nll": 2.3503409054085354
        },
        {
          "id": "final-013",
          "domain": "en",
          "delta_nll": 0.07294375213365051,
          "nll": 2.544624949335339
        },
        {
          "id": "final-046",
          "domain": "code",
          "delta_nll": 0.06624053336766729,
          "nll": 1.3426882499650208
        },
        {
          "id": "final-063",
          "domain": "ko",
          "delta_nll": 0.06309844567647593,
          "nll": 1.9939486240980704
        },
        {
          "id": "final-019",
          "domain": "en",
          "delta_nll": 0.06203505051631662,
          "nll": 1.9623728748317262
        },
        {
          "id": "final-047",
          "domain": "ko",
          "delta_nll": 0.05830257273553661,
          "nll": 1.5891902932813502
        },
        {
          "id": "final-024",
          "domain": "ko",
          "delta_nll": 0.056453347677868755,
          "nll": 2.1876879296649783
        },
        {
          "id": "final-124",
          "domain": "ko",
          "delta_nll": 0.05469006523076297,
          "nll": 2.1022687137347065
        },
        {
          "id": "final-030",
          "domain": "ko",
          "delta_nll": 0.05439239831284901,
          "nll": 1.5023053628017553
        },
        {
          "id": "final-073",
          "domain": "en",
          "delta_nll": 0.05397568013639561,
          "nll": 2.2551161028804056
        },
        {
          "id": "final-083",
          "domain": "code",
          "delta_nll": 0.053870922717069,
          "nll": 1.1102550361923647
        },
        {
          "id": "final-003",
          "domain": "ko",
          "delta_nll": 0.05277128865201296,
          "nll": 1.4713140320800644
        },
        {
          "id": "final-039",
          "domain": "ko",
          "delta_nll": 0.05266503023578243,
          "nll": 2.296586412909586
        },
        {
          "id": "final-032",
          "domain": "ko",
          "delta_nll": 0.05213395600446524,
          "nll": 1.8891258883176023
        },
        {
          "id": "final-126",
          "domain": "ko",
          "delta_nll": 0.051776652951370884,
          "nll": 2.4192633690632284
        },
        {
          "id": "final-101",
          "domain": "en",
          "delta_nll": 0.05169492799620634,
          "nll": 2.288882826523074
        },
        {
          "id": "final-034",
          "domain": "ko",
          "delta_nll": 0.05027131433247346,
          "nll": 2.032461391931003
        }
      ],
      "token_delta_tail": {
        "mean": 0.03653545246813151,
        "median": 0.0010466962703503668,
        "p95": 0.5288243308663364,
        "p99": 1.1602056407928452,
        "max": 9.285976648330688
      },
      "sampled_full_vocab_kl": null,
      "kl_positions": 0,
      "tasks": {
        "kmmlu": {
          "n": 500,
          "accuracy": 0.432,
          "failures": 0,
          "subject_scores": {
            "Accounting": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Agricultural-Sciences": {
              "n": 12,
              "accuracy": 0.25
            },
            "Aviation-Engineering-and-Maintenance": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Biology": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Chemical-Engineering": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Chemistry": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Civil-Engineering": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Computer-Science": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Construction": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Criminal-Law": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Ecology": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Economics": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Education": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Electrical-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Electronics-Engineering": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Energy-Management": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Environmental-Science": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Fashion": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Food-Processing": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Gas-Technology-and-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Geomatics": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Health": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Industrial-Engineer": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Information-Technology": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Interior-Architecture-and-Design": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Korean-History": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Law": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Machine-Design-and-Manufacturing": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Management": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Maritime-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Marketing": {
              "n": 11,
              "accuracy": 0.9090909090909091
            },
            "Materials-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Math": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Mechanical-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Nondestructive-Testing": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Patent": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Political-Science-and-Sociology": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Psychology": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Public-Safety": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Railway-and-Automotive-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Real-Estate": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Refrigerating-Machinery": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Social-Welfare": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Taxation": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Telecommunications-and-Wireless-Technology": {
              "n": 11,
              "accuracy": 0.7272727272727273
            }
          },
          "macro_accuracy": 0.43249158249158237,
          "delta_accuracy": {
            "delta": -0.056,
            "ci95": [
              -0.09,
              -0.023950000000000273
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "kmmlu-Accounting-18",
            "kmmlu-Accounting-52",
            "kmmlu-Accounting-35",
            "kmmlu-Agricultural-Sciences-999",
            "kmmlu-Agricultural-Sciences-660",
            "kmmlu-Agricultural-Sciences-226",
            "kmmlu-Aviation-Engineering-and-Maintenance-899",
            "kmmlu-Aviation-Engineering-and-Maintenance-808",
            "kmmlu-Aviation-Engineering-and-Maintenance-282",
            "kmmlu-Chemical-Engineering-584",
            "kmmlu-Chemistry-309",
            "kmmlu-Chemistry-486",
            "kmmlu-Civil-Engineering-837",
            "kmmlu-Computer-Science-66",
            "kmmlu-Construction-270",
            "kmmlu-Criminal-Law-61",
            "kmmlu-Ecology-998",
            "kmmlu-Ecology-577",
            "kmmlu-Ecology-248",
            "kmmlu-Economics-129",
            "kmmlu-Education-20",
            "kmmlu-Electrical-Engineering-135",
            "kmmlu-Electronics-Engineering-294",
            "kmmlu-Electronics-Engineering-810",
            "kmmlu-Electronics-Engineering-83",
            "kmmlu-Electronics-Engineering-78",
            "kmmlu-Energy-Management-645",
            "kmmlu-Energy-Management-431",
            "kmmlu-Energy-Management-155",
            "kmmlu-Environmental-Science-192",
            "kmmlu-Environmental-Science-342",
            "kmmlu-Fashion-331",
            "kmmlu-Food-Processing-148",
            "kmmlu-Gas-Technology-and-Engineering-348",
            "kmmlu-Industrial-Engineer-706",
            "kmmlu-Information-Technology-258",
            "kmmlu-Information-Technology-635",
            "kmmlu-Interior-Architecture-and-Design-748",
            "kmmlu-Interior-Architecture-and-Design-849",
            "kmmlu-Interior-Architecture-and-Design-560",
            "kmmlu-Law-755",
            "kmmlu-Machine-Design-and-Manufacturing-699",
            "kmmlu-Management-192",
            "kmmlu-Maritime-Engineering-115",
            "kmmlu-Marketing-501",
            "kmmlu-Math-246",
            "kmmlu-Patent-28",
            "kmmlu-Public-Safety-262",
            "kmmlu-Railway-and-Automotive-Engineering-685",
            "kmmlu-Railway-and-Automotive-Engineering-250",
            "kmmlu-Social-Welfare-19",
            "kmmlu-Taxation-95"
          ],
          "improvements": [
            "kmmlu-Accounting-24",
            "kmmlu-Aviation-Engineering-and-Maintenance-303",
            "kmmlu-Aviation-Engineering-and-Maintenance-380",
            "kmmlu-Chemical-Engineering-583",
            "kmmlu-Criminal-Law-117",
            "kmmlu-Electrical-Engineering-386",
            "kmmlu-Energy-Management-826",
            "kmmlu-Environmental-Science-318",
            "kmmlu-Interior-Architecture-and-Design-723",
            "kmmlu-Interior-Architecture-and-Design-561",
            "kmmlu-Interior-Architecture-and-Design-947",
            "kmmlu-Korean-History-89",
            "kmmlu-Korean-History-1",
            "kmmlu-Law-694",
            "kmmlu-Marketing-80",
            "kmmlu-Materials-Engineering-321",
            "kmmlu-Political-Science-and-Sociology-96",
            "kmmlu-Political-Science-and-Sociology-198",
            "kmmlu-Psychology-939",
            "kmmlu-Public-Safety-152",
            "kmmlu-Refrigerating-Machinery-473",
            "kmmlu-Refrigerating-Machinery-302",
            "kmmlu-Social-Welfare-606",
            "kmmlu-Telecommunications-and-Wireless-Technology-875"
          ]
        },
        "arc": {
          "n": 500,
          "accuracy": 0.524,
          "failures": 0,
          "accuracy_norm": 0.548,
          "delta_accuracy": {
            "delta": 0.002,
            "ci95": [
              -0.022,
              0.024
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "arc-Mercury_7025060",
            "arc-Mercury_SC_414271",
            "arc-Mercury_7211680",
            "arc-MCAS_2005_9_6",
            "arc-Mercury_7007420",
            "arc-Mercury_7093048",
            "arc-Mercury_SC_407689",
            "arc-Mercury_SC_406451",
            "arc-TIMSS_2007_8_pg113",
            "arc-Mercury_417138",
            "arc-Mercury_SC_409241",
            "arc-Mercury_SC_400839",
            "arc-Mercury_7236740",
            "arc-MCAS_1999_4_23",
            "arc-MCAS_2010_8_12005",
            "arc-Mercury_7212328",
            "arc-Mercury_7201740"
          ],
          "improvements": [
            "arc-Mercury_7143308",
            "arc-Mercury_SC_400594",
            "arc-MSA_2012_5_2",
            "arc-Mercury_SC_407613",
            "arc-Mercury_183033",
            "arc-Mercury_7082320",
            "arc-NYSEDREGENTS_2008_4_15",
            "arc-Mercury_7216248",
            "arc-Mercury_SC_405937",
            "arc-Mercury_SC_402164",
            "arc-AIMS_2009_4_5",
            "arc-Mercury_415303",
            "arc-Mercury_189070",
            "arc-Mercury_7185343",
            "arc-Mercury_7252245",
            "arc-Mercury_411732",
            "arc-Mercury_7159075",
            "arc-Mercury_SC_400708"
          ],
          "delta_accuracy_norm": {
            "delta": -0.01,
            "ci95": [
              -0.034,
              0.014
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          }
        },
        "ifeval": {
          "n": 541,
          "accuracy": 0.7966728280961183,
          "failures": 0,
          "instruction_strict": 0.8573141486810552,
          "truncated": 19,
          "delta_accuracy": {
            "delta": -0.03512014787430684,
            "ci95": [
              -0.06469500924214418,
              -0.0073937153419593345
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "ifeval-1122",
            "ifeval-1139",
            "ifeval-1219",
            "ifeval-1242",
            "ifeval-1265",
            "ifeval-1348",
            "ifeval-1392",
            "ifeval-1436",
            "ifeval-1609",
            "ifeval-1643",
            "ifeval-1670",
            "ifeval-1713",
            "ifeval-179",
            "ifeval-1823",
            "ifeval-1908",
            "ifeval-1967",
            "ifeval-1996",
            "ifeval-2041",
            "ifeval-2063",
            "ifeval-2069",
            "ifeval-2247",
            "ifeval-2250",
            "ifeval-2275",
            "ifeval-2589",
            "ifeval-2602",
            "ifeval-2605",
            "ifeval-2653",
            "ifeval-2716",
            "ifeval-2790",
            "ifeval-2832",
            "ifeval-2912",
            "ifeval-2929",
            "ifeval-296",
            "ifeval-3069",
            "ifeval-3098",
            "ifeval-3156",
            "ifeval-3276",
            "ifeval-3315",
            "ifeval-3445",
            "ifeval-3623"
          ],
          "improvements": [
            "ifeval-1092",
            "ifeval-1203",
            "ifeval-1629",
            "ifeval-2084",
            "ifeval-2180",
            "ifeval-2266",
            "ifeval-2299",
            "ifeval-2571",
            "ifeval-2662",
            "ifeval-2674",
            "ifeval-2941",
            "ifeval-3035",
            "ifeval-3195",
            "ifeval-3256",
            "ifeval-331",
            "ifeval-3351",
            "ifeval-3409",
            "ifeval-3415",
            "ifeval-3453",
            "ifeval-3534",
            "ifeval-3691"
          ]
        }
      },
      "speed": {
        "512": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.04959464073181152,
            "min": 0.04455304145812988,
            "max": 0.10157060623168945
          },
          "prefill_seconds": {
            "median": 0.048156779999999344,
            "min": 0.043370059999972455,
            "max": 0.10035488899995926
          },
          "decode_seconds": {
            "median": 5.650424885000007,
            "min": 4.513579437000033,
            "max": 6.6403984089999994
          },
          "seconds": {
            "median": 5.695060162999994,
            "min": 4.557910619999973,
            "max": 6.689486945999988
          },
          "prefill_tokens_per_second": {
            "median": 10631.940092340206,
            "min": 5101.893939618705,
            "max": 11805.379102549667
          },
          "decode_tokens_per_second": {
            "median": 45.129349595804726,
            "min": 38.40131032716173,
            "max": 56.49618081597046
          },
          "peak_device_vram_bytes": 30429831168
        },
        "4096": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.2177879810333252,
            "min": 0.21390390396118164,
            "max": 0.21947574615478516
          },
          "prefill_seconds": {
            "median": 0.2161841930000037,
            "min": 0.21227339899996878,
            "max": 0.21789716699998962
          },
          "decode_seconds": {
            "median": 5.081278778000012,
            "min": 4.358565473999988,
            "max": 7.047872522999967
          },
          "seconds": {
            "median": 5.300327787000015,
            "min": 4.576107018000016,
            "max": 7.261361922999981
          },
          "prefill_tokens_per_second": {
            "median": 18946.806161724922,
            "min": 18797.85798224809,
            "max": 19295.87041662532
          },
          "decode_tokens_per_second": {
            "median": 50.18421762333769,
            "min": 36.18113113820308,
            "max": 58.50548799166684
          },
          "peak_device_vram_bytes": 31086006272
        }
      },
      "speed_source": "dedicated serial benchmark",
      "runtime_memory_breakdown": {
        "source": "vLLM startup profile, rounded GiB",
        "weight_gib": 3.11,
        "profile_peak_activation_gib": 0.42,
        "non_torch_gib": 0.15,
        "reserved_kv_cache_gib": 21.77
      }
    },
    "vllm-R0": {
      "engine": "vllm",
      "candidate": "R0",
      "records": 4714,
      "failed_requests": [],
      "final_document_expected": 128,
      "final_document_present": 128,
      "final_document_failed": 0,
      "quality_complete": true,
      "nll": 1.9406551411408473,
      "ppl": 6.963311428395681,
      "scored_tokens": 130944,
      "domains": {
        "ko": {
          "documents": 64,
          "nll": 2.1986095356203292,
          "ppl": 9.012473260035641
        },
        "en": {
          "documents": 32,
          "nll": 2.2330683326369427,
          "ppl": 9.328444984466172
        },
        "code": {
          "documents": 32,
          "nll": 1.132333160685786,
          "ppl": 3.1028875965421494
        }
      },
      "delta_nll": {
        "delta": 0.08882180277252402,
        "ci95": [
          0.08290276643344494,
          0.09480845128102602
        ],
        "resampling_unit": "paired document",
        "replicates": 2000
      },
      "ppl_ratio_vs_bf16": {
        "ratio": 1.0928858897379137,
        "ci95": [
          1.0864361653490975,
          1.09944823705355
        ]
      },
      "document_delta_tail": {
        "mean": 0.08882180277252402,
        "median": 0.087994404574261,
        "p95": 0.14856161287850198,
        "p99": 0.1587154444139622,
        "max": 0.19832272298927292
      },
      "worst20_documents": [
        {
          "id": "final-049",
          "domain": "code",
          "delta_nll": 0.19832272298927292,
          "nll": 0.7246363137150335
        },
        {
          "id": "final-063",
          "domain": "ko",
          "delta_nll": 0.15986515114203179,
          "nll": 2.090715329563626
        },
        {
          "id": "final-003",
          "domain": "ko",
          "delta_nll": 0.1556069780751074,
          "nll": 1.5741497215031588
        },
        {
          "id": "final-103",
          "domain": "en",
          "delta_nll": 0.15501229391128007,
          "nll": 2.2303212552335223
        },
        {
          "id": "final-001",
          "domain": "code",
          "delta_nll": 0.1546165162903852,
          "nll": 0.35487132863211684
        },
        {
          "id": "final-085",
          "domain": "ko",
          "delta_nll": 0.14946690995447537,
          "nll": 2.7799039448369816
        },
        {
          "id": "final-035",
          "domain": "ko",
          "delta_nll": 0.14940944534518205,
          "nll": 2.2389908660766067
        },
        {
          "id": "final-018",
          "domain": "ko",
          "delta_nll": 0.14698706686895333,
          "nll": 2.44156141958132
        },
        {
          "id": "final-012",
          "domain": "code",
          "delta_nll": 0.14389871201080529,
          "nll": 0.4193171412541752
        },
        {
          "id": "final-033",
          "domain": "ko",
          "delta_nll": 0.13999413301145092,
          "nll": 2.486811311170567
        },
        {
          "id": "final-047",
          "domain": "ko",
          "delta_nll": 0.1387310207884297,
          "nll": 1.6696187413342434
        },
        {
          "id": "final-036",
          "domain": "ko",
          "delta_nll": 0.13660915660273554,
          "nll": 2.4098319480985144
        },
        {
          "id": "final-075",
          "domain": "ko",
          "delta_nll": 0.13436946684833534,
          "nll": 2.0693179257197922
        },
        {
          "id": "final-126",
          "domain": "ko",
          "delta_nll": 0.1289476009822346,
          "nll": 2.496434317094092
        },
        {
          "id": "final-052",
          "domain": "code",
          "delta_nll": 0.1285562151468034,
          "nll": 1.2281253555741551
        },
        {
          "id": "final-118",
          "domain": "ko",
          "delta_nll": 0.12769279741910605,
          "nll": 2.2424166882698624
        },
        {
          "id": "final-009",
          "domain": "ko",
          "delta_nll": 0.12623405137679639,
          "nll": 2.3612636813955876
        },
        {
          "id": "final-046",
          "domain": "code",
          "delta_nll": 0.12368547436031965,
          "nll": 1.400133190957673
        },
        {
          "id": "final-039",
          "domain": "ko",
          "delta_nll": 0.12295584540712193,
          "nll": 2.3668772280809254
        },
        {
          "id": "final-057",
          "domain": "ko",
          "delta_nll": 0.12157007951263825,
          "nll": 2.605094729994394
        }
      ],
      "token_delta_tail": {
        "mean": 0.08882180277252401,
        "median": 0.005889688618481159,
        "p95": 0.9199948191642747,
        "p99": 1.9303989338874754,
        "max": 11.345199823379517
      },
      "sampled_full_vocab_kl": null,
      "kl_positions": 0,
      "tasks": {
        "kmmlu": {
          "n": 500,
          "accuracy": 0.472,
          "failures": 0,
          "subject_scores": {
            "Accounting": {
              "n": 12,
              "accuracy": 0.3333333333333333
            },
            "Agricultural-Sciences": {
              "n": 12,
              "accuracy": 0.4166666666666667
            },
            "Aviation-Engineering-and-Maintenance": {
              "n": 12,
              "accuracy": 0.5
            },
            "Biology": {
              "n": 12,
              "accuracy": 0.25
            },
            "Chemical-Engineering": {
              "n": 12,
              "accuracy": 0.5
            },
            "Chemistry": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Civil-Engineering": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Computer-Science": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Construction": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Criminal-Law": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Ecology": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Economics": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Education": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Electrical-Engineering": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Electronics-Engineering": {
              "n": 11,
              "accuracy": 0.09090909090909091
            },
            "Energy-Management": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Environmental-Science": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Fashion": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Food-Processing": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Gas-Technology-and-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Geomatics": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Health": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Industrial-Engineer": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Information-Technology": {
              "n": 11,
              "accuracy": 0.9090909090909091
            },
            "Interior-Architecture-and-Design": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Korean-History": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Law": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Machine-Design-and-Manufacturing": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Management": {
              "n": 11,
              "accuracy": 0.7272727272727273
            },
            "Maritime-Engineering": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Marketing": {
              "n": 11,
              "accuracy": 0.9090909090909091
            },
            "Materials-Engineering": {
              "n": 11,
              "accuracy": 0.36363636363636365
            },
            "Math": {
              "n": 11,
              "accuracy": 0.18181818181818182
            },
            "Mechanical-Engineering": {
              "n": 11,
              "accuracy": 0.5454545454545454
            },
            "Nondestructive-Testing": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Patent": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Political-Science-and-Sociology": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Psychology": {
              "n": 11,
              "accuracy": 0.2727272727272727
            },
            "Public-Safety": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Railway-and-Automotive-Engineering": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Real-Estate": {
              "n": 11,
              "accuracy": 0.6363636363636364
            },
            "Refrigerating-Machinery": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Social-Welfare": {
              "n": 11,
              "accuracy": 0.8181818181818182
            },
            "Taxation": {
              "n": 11,
              "accuracy": 0.45454545454545453
            },
            "Telecommunications-and-Wireless-Technology": {
              "n": 11,
              "accuracy": 0.6363636363636364
            }
          },
          "macro_accuracy": 0.47272727272727266,
          "delta_accuracy": {
            "delta": -0.016,
            "ci95": [
              -0.058,
              0.024
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "kmmlu-Accounting-18",
            "kmmlu-Accounting-52",
            "kmmlu-Accounting-35",
            "kmmlu-Agricultural-Sciences-226",
            "kmmlu-Aviation-Engineering-and-Maintenance-899",
            "kmmlu-Biology-379",
            "kmmlu-Biology-731",
            "kmmlu-Chemistry-367",
            "kmmlu-Chemistry-309",
            "kmmlu-Chemistry-486",
            "kmmlu-Computer-Science-714",
            "kmmlu-Criminal-Law-73",
            "kmmlu-Ecology-577",
            "kmmlu-Ecology-248",
            "kmmlu-Economics-104",
            "kmmlu-Economics-129",
            "kmmlu-Economics-106",
            "kmmlu-Education-20",
            "kmmlu-Electrical-Engineering-135",
            "kmmlu-Electronics-Engineering-294",
            "kmmlu-Electronics-Engineering-111",
            "kmmlu-Electronics-Engineering-810",
            "kmmlu-Electronics-Engineering-83",
            "kmmlu-Electronics-Engineering-78",
            "kmmlu-Energy-Management-645",
            "kmmlu-Energy-Management-155",
            "kmmlu-Environmental-Science-192",
            "kmmlu-Environmental-Science-342",
            "kmmlu-Fashion-331",
            "kmmlu-Fashion-419",
            "kmmlu-Fashion-249",
            "kmmlu-Food-Processing-225",
            "kmmlu-Gas-Technology-and-Engineering-361",
            "kmmlu-Gas-Technology-and-Engineering-311",
            "kmmlu-Geomatics-272",
            "kmmlu-Health-91",
            "kmmlu-Health-62",
            "kmmlu-Information-Technology-635",
            "kmmlu-Interior-Architecture-and-Design-849",
            "kmmlu-Korean-History-93",
            "kmmlu-Korean-History-28",
            "kmmlu-Machine-Design-and-Manufacturing-690",
            "kmmlu-Machine-Design-and-Manufacturing-815",
            "kmmlu-Machine-Design-and-Manufacturing-345",
            "kmmlu-Management-192",
            "kmmlu-Management-80",
            "kmmlu-Maritime-Engineering-145",
            "kmmlu-Marketing-521",
            "kmmlu-Nondestructive-Testing-500",
            "kmmlu-Nondestructive-Testing-286",
            "kmmlu-Patent-99",
            "kmmlu-Political-Science-and-Sociology-123",
            "kmmlu-Psychology-458",
            "kmmlu-Public-Safety-941",
            "kmmlu-Public-Safety-262",
            "kmmlu-Railway-and-Automotive-Engineering-685",
            "kmmlu-Railway-and-Automotive-Engineering-801",
            "kmmlu-Refrigerating-Machinery-797",
            "kmmlu-Refrigerating-Machinery-247",
            "kmmlu-Social-Welfare-952"
          ],
          "improvements": [
            "kmmlu-Aviation-Engineering-and-Maintenance-81",
            "kmmlu-Chemical-Engineering-837",
            "kmmlu-Civil-Engineering-594",
            "kmmlu-Civil-Engineering-924",
            "kmmlu-Civil-Engineering-102",
            "kmmlu-Computer-Science-923",
            "kmmlu-Construction-284",
            "kmmlu-Criminal-Law-40",
            "kmmlu-Ecology-737",
            "kmmlu-Ecology-254",
            "kmmlu-Economics-74",
            "kmmlu-Electrical-Engineering-474",
            "kmmlu-Electrical-Engineering-118",
            "kmmlu-Energy-Management-826",
            "kmmlu-Environmental-Science-318",
            "kmmlu-Environmental-Science-147",
            "kmmlu-Food-Processing-85",
            "kmmlu-Food-Processing-919",
            "kmmlu-Geomatics-15",
            "kmmlu-Geomatics-543",
            "kmmlu-Geomatics-900",
            "kmmlu-Health-40",
            "kmmlu-Health-58",
            "kmmlu-Industrial-Engineer-927",
            "kmmlu-Interior-Architecture-and-Design-618",
            "kmmlu-Interior-Architecture-and-Design-561",
            "kmmlu-Korean-History-89",
            "kmmlu-Korean-History-64",
            "kmmlu-Law-11",
            "kmmlu-Law-277",
            "kmmlu-Management-162",
            "kmmlu-Management-229",
            "kmmlu-Marketing-80",
            "kmmlu-Materials-Engineering-321",
            "kmmlu-Mechanical-Engineering-629",
            "kmmlu-Mechanical-Engineering-605",
            "kmmlu-Patent-66",
            "kmmlu-Patent-48",
            "kmmlu-Political-Science-and-Sociology-96",
            "kmmlu-Political-Science-and-Sociology-198",
            "kmmlu-Political-Science-and-Sociology-213",
            "kmmlu-Public-Safety-651",
            "kmmlu-Public-Safety-152",
            "kmmlu-Railway-and-Automotive-Engineering-834",
            "kmmlu-Railway-and-Automotive-Engineering-177",
            "kmmlu-Real-Estate-46",
            "kmmlu-Real-Estate-172",
            "kmmlu-Real-Estate-155",
            "kmmlu-Refrigerating-Machinery-473",
            "kmmlu-Refrigerating-Machinery-302",
            "kmmlu-Social-Welfare-706",
            "kmmlu-Taxation-185"
          ]
        },
        "arc": {
          "n": 500,
          "accuracy": 0.5,
          "failures": 0,
          "accuracy_norm": 0.548,
          "delta_accuracy": {
            "delta": -0.022,
            "ci95": [
              -0.05,
              0.004
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "arc-Mercury_7025060",
            "arc-Mercury_7162400",
            "arc-Mercury_SC_414271",
            "arc-Mercury_7086608",
            "arc-MCAS_2004_8_14",
            "arc-Mercury_7007420",
            "arc-Mercury_7093048",
            "arc-Mercury_SC_407689",
            "arc-Mercury_SC_406451",
            "arc-Mercury_415084",
            "arc-Mercury_7233765",
            "arc-TIMSS_2007_8_pg113",
            "arc-Mercury_417138",
            "arc-Mercury_SC_400839",
            "arc-Mercury_7136885",
            "arc-MCAS_2004_9_2",
            "arc-Mercury_7132405",
            "arc-Mercury_175560",
            "arc-Mercury_7042945",
            "arc-Mercury_187198",
            "arc-Mercury_7227710",
            "arc-Mercury_184730",
            "arc-Mercury_7012740",
            "arc-MDSA_2007_8_42",
            "arc-Mercury_SC_407400",
            "arc-Mercury_403907",
            "arc-Mercury_402364"
          ],
          "improvements": [
            "arc-ACTAAP_2011_5_1",
            "arc-Mercury_7143308",
            "arc-Mercury_SC_407613",
            "arc-Mercury_7206395",
            "arc-NYSEDREGENTS_2008_4_15",
            "arc-MDSA_2008_8_27",
            "arc-Mercury_SC_LBS10337",
            "arc-NCEOGA_2013_8_7",
            "arc-Mercury_417118",
            "arc-Mercury_7268240",
            "arc-Mercury_7283833",
            "arc-AIMS_2009_4_5",
            "arc-MEA_2016_8_15",
            "arc-Mercury_189070",
            "arc-NCEOGA_2013_8_28",
            "arc-Mercury_7056298"
          ],
          "delta_accuracy_norm": {
            "delta": -0.01,
            "ci95": [
              -0.034,
              0.014
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          }
        },
        "ifeval": {
          "n": 541,
          "accuracy": 0.7652495378927912,
          "failures": 0,
          "instruction_strict": 0.8309352517985612,
          "truncated": 45,
          "delta_accuracy": {
            "delta": -0.066543438077634,
            "ci95": [
              -0.09611829944547134,
              -0.036968576709796676
            ],
            "resampling_unit": "paired item",
            "replicates": 2000
          },
          "regressions": [
            "ifeval-1219",
            "ifeval-1233",
            "ifeval-1248",
            "ifeval-1265",
            "ifeval-1268",
            "ifeval-1342",
            "ifeval-1348",
            "ifeval-1367",
            "ifeval-1381",
            "ifeval-1392",
            "ifeval-1436",
            "ifeval-1466",
            "ifeval-1571",
            "ifeval-1609",
            "ifeval-1670",
            "ifeval-1813",
            "ifeval-1825",
            "ifeval-1837",
            "ifeval-1908",
            "ifeval-1967",
            "ifeval-2035",
            "ifeval-2063",
            "ifeval-2243",
            "ifeval-2246",
            "ifeval-2247",
            "ifeval-2273",
            "ifeval-2355",
            "ifeval-2386",
            "ifeval-24",
            "ifeval-2416",
            "ifeval-2439",
            "ifeval-2596",
            "ifeval-2605",
            "ifeval-2616",
            "ifeval-2637",
            "ifeval-2677",
            "ifeval-2716",
            "ifeval-279",
            "ifeval-2908",
            "ifeval-2909",
            "ifeval-2912",
            "ifeval-2969",
            "ifeval-30",
            "ifeval-3025",
            "ifeval-3098",
            "ifeval-3186",
            "ifeval-3223",
            "ifeval-3243",
            "ifeval-3335",
            "ifeval-3350",
            "ifeval-3445",
            "ifeval-3479",
            "ifeval-3540",
            "ifeval-3748"
          ],
          "improvements": [
            "ifeval-1092",
            "ifeval-1203",
            "ifeval-1645",
            "ifeval-2078",
            "ifeval-2084",
            "ifeval-2299",
            "ifeval-2350",
            "ifeval-2674",
            "ifeval-2941",
            "ifeval-3035",
            "ifeval-3256",
            "ifeval-3294",
            "ifeval-331",
            "ifeval-3329",
            "ifeval-3409",
            "ifeval-3453",
            "ifeval-3557",
            "ifeval-3691"
          ]
        }
      },
      "speed": {
        "512": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.04797053337097168,
            "min": 0.04520893096923828,
            "max": 0.09441709518432617
          },
          "prefill_seconds": {
            "median": 0.04673513299999854,
            "min": 0.04410053699999139,
            "max": 0.09323427500001458
          },
          "decode_seconds": {
            "median": 5.031832434000023,
            "min": 4.672067717999994,
            "max": 5.736560865000001
          },
          "seconds": {
            "median": 5.076876428999981,
            "min": 4.71994102299999,
            "max": 5.784405703999994
          },
          "prefill_tokens_per_second": {
            "median": 10955.355577997734,
            "min": 5491.542675694319,
            "max": 11609.835952793499
          },
          "decode_tokens_per_second": {
            "median": 50.677363235899605,
            "min": 44.451720464749215,
            "max": 54.57968835031349
          },
          "peak_device_vram_bytes": 30514843648
        },
        "4096": {
          "measured": 5,
          "failed": 0,
          "ttft_seconds": {
            "median": 0.19234180450439453,
            "min": 0.19208455085754395,
            "max": 0.24526643753051758
          },
          "prefill_seconds": {
            "median": 0.19094724900000415,
            "min": 0.19074504299999262,
            "max": 0.24380702399997745
          },
          "decode_seconds": {
            "median": 5.615284588999998,
            "min": 4.693565898000003,
            "max": 7.128385011999995
          },
          "seconds": {
            "median": 5.812666755999999,
            "min": 4.886212628999999,
            "max": 7.3204510170000106
          },
          "prefill_tokens_per_second": {
            "median": 21450.95057117011,
            "min": 16800.172254267698,
            "max": 21473.690406728725
          },
          "decode_tokens_per_second": {
            "median": 45.41176782019731,
            "min": 35.77247855871006,
            "max": 54.32969421152886
          },
          "peak_device_vram_bytes": 31203332096
        }
      },
      "speed_source": "dedicated serial benchmark",
      "runtime_memory_breakdown": {
        "source": "vLLM startup profile, rounded GiB",
        "weight_gib": 3.11,
        "profile_peak_activation_gib": 0.42,
        "non_torch_gib": 0.15,
        "reserved_kv_cache_gib": 21.77
      }
    }
  },
  "sizes": {
    "native-B0": {
      "file_bytes": 8424393504,
      "scope": "text GGUF including metadata",
      "text_tensor_payload_bytes": 8413425664,
      "original_text_elements": 4205751296,
      "effective_payload_bpw": 16.003657985201034
    },
    "native-K0": {
      "file_bytes": 3299478304,
      "scope": "text GGUF including metadata",
      "text_tensor_payload_bytes": 3288510464,
      "original_text_elements": 4205751296,
      "effective_payload_bpw": 6.255263771070119
    },
    "native-K1": {
      "file_bytes": 3299478624,
      "scope": "text GGUF including metadata",
      "text_tensor_payload_bytes": 3288510464,
      "original_text_elements": 4205751296,
      "effective_payload_bpw": 6.255263771070119
    },
    "native-KM": {
      "file_bytes": 3464383584,
      "scope": "text GGUF including metadata",
      "text_tensor_payload_bytes": 3453415424,
      "original_text_elements": 4205751296,
      "effective_payload_bpw": 6.5689389237722535
    },
    "native-I4": {
      "file_bytes": 3188067424,
      "scope": "text GGUF including metadata",
      "text_tensor_payload_bytes": 3177099264,
      "original_text_elements": 4205751296,
      "effective_payload_bpw": 6.043342157719447
    },
    "hf-B0": {
      "safetensors_file_bytes": 9319828096,
      "text_tensor_payload_bytes": 8411510272,
      "whole_directory_bytes": 9342911475,
      "original_text_elements": 4205751296,
      "effective_payload_bpw": 16.000014608567096,
      "scope": "text tensor payload excludes visual and MTP; files may include vision"
    },
    "hf-R0": {
      "safetensors_file_bytes": 3786643328,
      "text_tensor_payload_bytes": 3119473792,
      "whole_directory_bytes": 3806655612,
      "original_text_elements": 4205751296,
      "effective_payload_bpw": 5.933729452746033,
      "scope": "text tensor payload excludes visual and MTP; files may include vision"
    },
    "hf-G0": {
      "safetensors_file_bytes": 3786643328,
      "text_tensor_payload_bytes": 3119473792,
      "whole_directory_bytes": 3806655682,
      "original_text_elements": 4205751296,
      "effective_payload_bpw": 5.933729452746033,
      "scope": "text tensor payload excludes visual and MTP; files may include vision"
    },
    "hf-A0": {
      "safetensors_file_bytes": 3786643328,
      "text_tensor_payload_bytes": 3119473792,
      "whole_directory_bytes": 3806656527,
      "original_text_elements": 4205751296,
      "effective_payload_bpw": 5.933729452746033,
      "scope": "text tensor payload excludes visual and MTP; files may include vision"
    }
  },
  "task_manifest": {
    "seed": 20260921,
    "dataset_revisions": {
      "HAERAE-HUB/KMMLU": "d61b3f19e552c576bf5960dd24289763edc36a88",
      "allenai/ai2_arc": "210d026faf9955653af8916fad021475a3f00453",
      "google/IFEval": "966cd89545d6b6acfd7638bc708b98261ca58e84"
    },
    "kmmlu_configs": [
      "Accounting",
      "Agricultural-Sciences",
      "Aviation-Engineering-and-Maintenance",
      "Biology",
      "Chemical-Engineering",
      "Chemistry",
      "Civil-Engineering",
      "Computer-Science",
      "Construction",
      "Criminal-Law",
      "Ecology",
      "Economics",
      "Education",
      "Electrical-Engineering",
      "Electronics-Engineering",
      "Energy-Management",
      "Environmental-Science",
      "Fashion",
      "Food-Processing",
      "Gas-Technology-and-Engineering",
      "Geomatics",
      "Health",
      "Industrial-Engineer",
      "Information-Technology",
      "Interior-Architecture-and-Design",
      "Korean-History",
      "Law",
      "Machine-Design-and-Manufacturing",
      "Management",
      "Maritime-Engineering",
      "Marketing",
      "Materials-Engineering",
      "Math",
      "Mechanical-Engineering",
      "Nondestructive-Testing",
      "Patent",
      "Political-Science-and-Sociology",
      "Psychology",
      "Public-Safety",
      "Railway-and-Automotive-Engineering",
      "Real-Estate",
      "Refrigerating-Machinery",
      "Social-Welfare",
      "Taxation",
      "Telecommunications-and-Wireless-Technology"
    ],
    "counts": {
      "kmmlu": 500,
      "arc": 500,
      "ifeval": 541
    },
    "files": {
      "task-samples.jsonl": "149d8a55424bfa9a153dd28bcec55ba5901cc336f55d110349feea93bca94069",
      "task-requests.jsonl": "39c20c909b7144c7834a104d3e16506535e147503cb8acce4d3ab52271ee35e8",
      "document-requests.jsonl": "066f4c6fb958b272c683a181d03896a5036a5a3052e7fba89cd3377f168c3b94",
      "perf-requests.jsonl": "eadfe7314158d04b1e2b32c8b47691205fe6798e11e6b9d617d7887d10deb87f"
    },
    "llm_eval_version": "0.4.13",
    "mc_template": "lm-eval zero-shot raw completion templates, no chat wrapper",
    "ifeval": "original tokenizer chat template, enable_thinking=False, greedy, max_new_tokens=2048",
    "kl": "32 fixed uniformly sampled target positions per document; exact full-vocabulary KL(BF16||candidate)",
    "nll": "each independent <=1024 token document; score positions 1..end; no added BOS, no overlap, fresh state",
    "created_unix": 1790017507.9748929
  },
  "baseline_agreement": {
    "scope": "two fixed validation documents; not final evaluation",
    "documents": [
      {
        "id": "validation-000",
        "kl_hf_bf16_to_native_bf16_mean": 0.00032668747771120435,
        "kl_p95": 0.0010869226979381532,
        "kl_max": 0.0030274173200839924,
        "argmax_agreement": 1.0,
        "positions": 32,
        "self_kl_max": 0.0
      },
      {
        "id": "validation-001",
        "kl_hf_bf16_to_native_bf16_mean": 0.00026648011406643114,
        "kl_p95": 0.0005722330255313285,
        "kl_max": 0.0006650254597962712,
        "argmax_agreement": 0.96875,
        "positions": 32,
        "self_kl_max": 0.0
      }
    ],
    "all_finite": true,
    "mean_kl": 0.00029658379588881774,
    "gate_passed": true,
    "interpretation": "Engine numerical difference is measured; compare quantization deltas against each engine BF16. This is not proof of identical kernels."
  },
  "quantization_costs": {
    "R0": {
      "candidate": "R0",
      "pilot": false,
      "quantization_seconds": 0.47693753242492676,
      "total_seconds": 61.78158926963806,
      "peak_allocated_bytes": 9346222080,
      "peak_reserved_bytes": 9835642880,
      "payload_bytes": 3786643328,
      "packed_linear_count": 200,
      "path": "/models/huggingface/Qwen3.5-4B-QuantStudy-20260922-R0",
      "completed_at_unix": 1790015632.2326403,
      "reload_verified_at_creation": false,
      "reload_checks_now": {
        "hf": true,
        "vllm": true
      }
    },
    "G0": {
      "candidate": "G0",
      "pilot": false,
      "quantization_seconds": 347.3705756664276,
      "total_seconds": 407.3032581806183,
      "peak_allocated_bytes": 10689680384,
      "peak_reserved_bytes": 11249123328,
      "payload_bytes": 3786643328,
      "packed_linear_count": 200,
      "path": "/models/huggingface/Qwen3.5-4B-QuantStudy-20260922-G0",
      "completed_at_unix": 1790016046.3189902,
      "reload_verified_at_creation": false,
      "reload_checks_now": {
        "hf": true,
        "vllm": true
      }
    },
    "A0": {
      "candidate": "A0",
      "pilot": false,
      "quantization_seconds": 230.36617827415466,
      "total_seconds": 291.3701422214508,
      "peak_allocated_bytes": 14561832448,
      "peak_reserved_bytes": 15319695360,
      "payload_bytes": 3786643328,
      "packed_linear_count": 200,
      "path": "/models/huggingface/Qwen3.5-4B-QuantStudy-20260922-A0",
      "completed_at_unix": 1790017417.474299,
      "reload_verified_at_creation": false,
      "reload_checks_now": {
        "hf": true,
        "vllm": true
      }
    },
    "K0": {
      "process_wall_seconds": 16.343690872192383,
      "scope": "native process including model IO; GPU peak/CPU peak not separately sampled"
    },
    "K1": {
      "process_wall_seconds": 23.947975635528564,
      "scope": "native process including model IO; GPU peak/CPU peak not separately sampled"
    },
    "KM": {
      "process_wall_seconds": 21.803329467773438,
      "scope": "native process including model IO; GPU peak/CPU peak not separately sampled"
    },
    "I4": {
      "process_wall_seconds": 40.300732135772705,
      "scope": "native process including model IO; GPU peak/CPU peak not separately sampled"
    },
    "shared_imatrix": {
      "process_wall_seconds": 64.55161738395691,
      "used_by": [
        "K1",
        "KM",
        "I4"
      ]
    }
  },
  "reboot_recovery": {
    "incident": "results/crash-20260922/INCIDENT.md",
    "bugcheck": "0xD1",
    "reported_related_driver": "NETIO.SYS",
    "root_cause_confirmed": false,
    "artifact_revalidation": {
      "ok": true,
      "time": 1790043521.4152524,
      "files": [
        {
          "path": "/workspace/quantization-study/run\\data\\calibration-documents.jsonl",
          "bytes": 1409817,
          "digest": "bc659614d8b2eeda7c44fcff5964ec8b9abcd5a137ffc690e82fbefc05c30795",
          "ok": true
        },
        {
          "path": "/workspace/quantization-study/run\\data\\validation-documents.jsonl",
          "bytes": 358229,
          "digest": "e72e0a1781759688ee332f787133d2346fdd3e4d698bdc47c342974b138f2f50",
          "ok": true
        },
        {
          "path": "/workspace/quantization-study/run\\data\\final-documents.jsonl",
          "bytes": 1400177,
          "digest": "faa9b0dfc1a46b365e372bb3e55232c888c6aca16831afb12ed9e5ba6fc94426",
          "ok": true
        },
        {
          "path": "/workspace/quantization-study/run\\data\\calibration.txt",
          "bytes": 520312,
          "digest": "d6e3c3a988fea455200d9b5608b98e650d3bf322c97fe2e58834779425ba5028",
          "ok": true
        },
        {
          "path": "/workspace/quantization-study/run\\data\\calibration-windows.jsonl",
          "bytes": 1207117,
          "digest": "1af35598ac67eb4ca3fa34f6764428feb88d2f9822d3bffdd9bc8a0ef44ecd58",
          "ok": true
        },
        {
          "path": "/workspace/quantization-study/run\\data\\calibration-tokenization.json",
          "bytes": 293,
          "digest": "3d3502fdbe3c80febd9a1f4534c755af61da086d3b2ef4c9f2318bb686cbd2c1",
          "ok": true
        },
        {
          "path": "/workspace/quantization-study/run\\data\\task-samples.jsonl",
          "bytes": 1385218,
          "digest": "149d8a55424bfa9a153dd28bcec55ba5901cc336f55d110349feea93bca94069",
          "ok": true
        },
        {
          "path": "/workspace/quantization-study/run\\data\\task-requests.jsonl",
          "bytes": 2594889,
          "digest": "39c20c909b7144c7834a104d3e16506535e147503cb8acce4d3ab52271ee35e8",
          "ok": true
        },
        {
          "path": "/workspace/quantization-study/run\\data\\document-requests.jsonl",
          "bytes": 1064042,
          "digest": "066f4c6fb958b272c683a181d03896a5036a5a3052e7fba89cd3377f168c3b94",
          "ok": true
        },
        {
          "path": "/workspace/quantization-study/run\\data\\perf-requests.jsonl",
          "bytes": 200749,
          "digest": "eadfe7314158d04b1e2b32c8b47691205fe6798e11e6b9d617d7887d10deb87f",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B\\.gitattributes",
          "bytes": 1570,
          "digest": "52373fe24473b1aa44333d318f578ae6bf04b49b",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B\\LICENSE",
          "bytes": 11544,
          "digest": "f938136e3adacfd92be087f6e113b5d6d97f678f",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B\\README.md",
          "bytes": 77661,
          "digest": "7950a3aadf378cd13758097bc52f0ed849a59007",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B\\chat_template.jinja",
          "bytes": 7756,
          "digest": "a585dec894e63da457d9440ec6aa7caa16d20860",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B\\config.json",
          "bytes": 3161,
          "digest": "557d961b205319c6a7da5f757f565b69b3967b7d",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B\\merges.txt",
          "bytes": 3353259,
          "digest": "a494e019ca1502219fd0128658b979e5f05ae8e8",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B\\model.safetensors-00001-of-00002.safetensors",
          "bytes": 5329398688,
          "digest": "26a93f066e1916adb13453dae5a0c707c0fbc71299ed98779571a907b8e74c61",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B\\model.safetensors-00002-of-00002.safetensors",
          "bytes": 3990429408,
          "digest": "cb544bd9bfae93dc59b0f22b292f5933573854a7f9b97835c67060d7d910e188",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B\\model.safetensors.index.json",
          "bytes": 76196,
          "digest": "fddda6039f7c1d17260c9e923b8a72fd025d9a86",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B\\preprocessor_config.json",
          "bytes": 390,
          "digest": "2ea84a437d448ff71b08df68fdd949d5cc4ebb64",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B\\tokenizer.json",
          "bytes": 12807982,
          "digest": "5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B\\tokenizer_config.json",
          "bytes": 16710,
          "digest": "eda48d3e75a8e59a8479ee4ec8b37f76e711d9c1",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B\\video_preprocessor_config.json",
          "bytes": 385,
          "digest": "3ba673a5ad7d4d13f54155ecd38b2a94a6dac8fe",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B\\vocab.json",
          "bytes": 6722759,
          "digest": "0aa0ce0658d60ac4a5d609f4eadb0e8e43514176",
          "ok": true
        },
        {
          "path": "/models\\gguf\\Qwen3.5-4B-QuantStudy-20260922\\Qwen3.5-4B-BF16.gguf",
          "bytes": 8424393504,
          "digest": "4f851fbf3f5efe9fe17d332aa7633e06f0cd6b8ca29cd74f0a78c55461734002",
          "ok": true
        },
        {
          "path": "/models\\gguf\\Qwen3.5-4B-QuantStudy-20260922\\Qwen3.5-4B-I4-IQ4_XS.gguf",
          "bytes": 3188067424,
          "digest": "d2f3dd7a1e6d919c32bb993aeca9bd994226899c462122bfbbca9a88d79a0dec",
          "ok": true
        },
        {
          "path": "/models\\gguf\\Qwen3.5-4B-QuantStudy-20260922\\Qwen3.5-4B-K0-Q4_K.gguf",
          "bytes": 3299478304,
          "digest": "e2065959661a406df842b9f6ee9f16b3054c236834a43c6e6b92c4c6100e6916",
          "ok": true
        },
        {
          "path": "/models\\gguf\\Qwen3.5-4B-QuantStudy-20260922\\Qwen3.5-4B-K1-Q4_K.gguf",
          "bytes": 3299478624,
          "digest": "fb879775d28280a13b6f983218637d7cde5916b8d49610e88ecacd1bfdbf3b17",
          "ok": true
        },
        {
          "path": "/models\\gguf\\Qwen3.5-4B-QuantStudy-20260922\\Qwen3.5-4B-KM-Q4_K_M.gguf",
          "bytes": 3464383584,
          "digest": "9826160dc86b8dce7e329d42323a6dfdbe8afda6d668bbcbcd0bae549772127a",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-R0\\chat_template.jinja",
          "bytes": 7756,
          "digest": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-R0\\config.json",
          "bytes": 13618,
          "digest": "4b6cd0d68cb8b673b612f8763fa0e7248674184c477904d2f3de182dbb9192b0",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-R0\\generation_config.json",
          "bytes": 116,
          "digest": "4a4505ef94dbb317b8c508ff53777cb0dfed044027864a1c05822c07217251ef",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-R0\\model.safetensors",
          "bytes": 3786643328,
          "digest": "c9caeccee9284be3a9d1d5c3dde923aea6b33ff42d1980001f344966830c6454",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-R0\\recipe.yaml",
          "bytes": 346,
          "digest": "4dc9a16dfdfdd18de75049d49d353d043f2e72d22cf1e66508466b42f0e4e323",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-R0\\tokenizer.json",
          "bytes": 19989325,
          "digest": "06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-R0\\tokenizer_config.json",
          "bytes": 1123,
          "digest": "bee8eba30f0eb4af73c0fe2cd06d0f89b657d7819941c438157ec42f7c80ea87",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-G0\\chat_template.jinja",
          "bytes": 7756,
          "digest": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-G0\\config.json",
          "bytes": 13618,
          "digest": "4b6cd0d68cb8b673b612f8763fa0e7248674184c477904d2f3de182dbb9192b0",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-G0\\generation_config.json",
          "bytes": 116,
          "digest": "4a4505ef94dbb317b8c508ff53777cb0dfed044027864a1c05822c07217251ef",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-G0\\model.safetensors",
          "bytes": 3786643328,
          "digest": "7b23eaf13cbe94ee041f052928d9eb7f25bb6810ad83093a5b4eb7a56829d6d4",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-G0\\recipe.yaml",
          "bytes": 416,
          "digest": "c05e0f30ae6e6bfdb12546f220587124ba7e77d9e26f6b685e1601328077f56a",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-G0\\tokenizer.json",
          "bytes": 19989325,
          "digest": "06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-G0\\tokenizer_config.json",
          "bytes": 1123,
          "digest": "bee8eba30f0eb4af73c0fe2cd06d0f89b657d7819941c438157ec42f7c80ea87",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-A0\\chat_template.jinja",
          "bytes": 7756,
          "digest": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-A0\\config.json",
          "bytes": 13618,
          "digest": "4b6cd0d68cb8b673b612f8763fa0e7248674184c477904d2f3de182dbb9192b0",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-A0\\generation_config.json",
          "bytes": 116,
          "digest": "4a4505ef94dbb317b8c508ff53777cb0dfed044027864a1c05822c07217251ef",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-A0\\model.safetensors",
          "bytes": 3786643328,
          "digest": "03d7ff5dba4a48ff1c689fd540ba6c03928b047a4ab9680bb5f3c3994bd82dca",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-A0\\recipe.yaml",
          "bytes": 1261,
          "digest": "88fef58471d11265dee9d59b0ab98a73335c08f17ca33a2b529bc217ab9620b3",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-A0\\tokenizer.json",
          "bytes": 19989325,
          "digest": "06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523",
          "ok": true
        },
        {
          "path": "/models\\huggingface\\Qwen3.5-4B-QuantStudy-20260922-A0\\tokenizer_config.json",
          "bytes": 1123,
          "digest": "bee8eba30f0eb4af73c0fe2cd06d0f89b657d7819941c438157ec42f7c80ea87",
          "ok": true
        }
      ],
      "awq_saved_requests": 4582,
      "awq_pending_requests": 132
    },
    "validation_replay": {
      "ok": true,
      "time": 1790043578.3605447,
      "documents": [
        {
          "id": "validation-000",
          "max_abs_token_nll_change": 0.0,
          "result": {
            "id": "validation-000",
            "nll_sum": 1390.3210441801912,
            "scored_tokens": 1023,
            "token_nll": [
              8.896589279174805,
              8.10812759399414,
              15.234284400939941,
              4.063816070556641,
              12.620002746582031,
              4.530691623687744,
              1.2930597066879272,
              0.35175272822380066,
              0.1414705514907837,
              2.2816872596740723,
              2.135037899017334,
              7.231238842010498,
              10.643446922302246,
              4.155224323272705,
              1.5417677164077759,
              4.567447185516357,
              0.31904101371765137,
              0.6470569968223572,
              1.9187898635864258,
              2.2457971572875977,
              0.9873025417327881,
              0.780295193195343,
              2.6049716472625732,
              0.13760657608509064,
              2.043778896331787,
              0.7781569361686707,
              3.7662031650543213,
              2.713412284851074,
              0.23097831010818481,
              6.861329078674316,
              2.480008363723755,
              0.8723762631416321,
              7.548976898193359,
              0.045730166137218475,
              6.295568943023682,
              1.7294002771377563,
              3.3896212577819824,
              1.2617295980453491,
              2.166970729827881,
              1.2645686864852905,
              1.8690170049667358,
              1.142823576927185,
              3.6533286571502686,
              0.9696319103240967,
              6.805932998657227,
              0.44700637459754944,
              1.6254490613937378,
              2.930039167404175,
              3.7419345378875732,
              1.2759637832641602,
              0.557578980922699,
              1.0683296918869019,
              0.7876671552658081,
              4.050053119659424,
              1.163008213043213,
              1.332289218902588,
              5.599356651306152,
              0.10200141370296478,
              0.7086512446403503,
              7.168918132781982,
              1.7046775817871094,
              10.218987464904785,
              7.706588268280029,
              0.16714976727962494,
              4.285050868988037,
              0.006876257248222828,
              6.07305908203125,
              0.065034881234169,
              7.29325008392334,
              4.127369403839111,
              2.7572555541992188,
              0.8364105224609375,
              4.476608753204346,
              1.7050001621246338,
              0.005860882345587015,
              5.669244289398193,
              0.1138870045542717,
              0.05124836787581444,
              0.33780398964881897,
              5.607630729675293,
              0.6737068891525269,
              3.84279727935791,
              4.671473026275635,
              0.006709665060043335,
              4.875365257263184,
              3.6566967964172363,
              8.060627937316895,
              3.104261875152588,
              3.0225369930267334,
              0.2533526122570038,
              0.13062703609466553,
              0.05253320187330246,
              8.621944427490234,
              0.346645712852478,
              0.5980298519134521,
              0.08677282929420471,
              0.00010263393050990999,
              0.11037678271532059,
              0.019280055537819862,
              0.001278298324905336,
              0.09665441513061523,
              0.07735737413167953,
              0.004121619276702404,
              0.16229917109012604,
              0.11573243886232376,
              7.676764488220215,
              0.019920174032449722,
              0.362432599067688,
              0.007550511043518782,
              0.0008995062671601772,
              0.08015387505292892,
              0.021290449425578117,
              0.00517433974891901,
              0.02617933228611946,
              0.4574211835861206,
              0.006322027184069157,
              0.20277181267738342,
              0.20573385059833527,
              5.205392837524414,
              0.007222376763820648,
              0.06693047285079956,
              0.0006850994541309774,
              2.3245540432981215e-05,
              0.005636157002300024,
              0.030755937099456787,
              0.0010303908493369818,
              0.022629013285040855,
              0.12353207170963287,
              0.005144335329532623,
              1.4072444438934326,
              1.9218240976333618,
              5.203754425048828,
              14.096646308898926,
              0.1556299328804016,
              1.6323039531707764,
              1.7160873413085938,
              5.02338981628418,
              2.0788657665252686,
              3.962313175201416,
              1.6145744323730469,
              0.19444392621517181,
              0.0822046622633934,
              0.0967492163181305,
              5.033435821533203,
              6.040292263031006,
              4.726674556732178,
              3.2992899417877197,
              0.11521605402231216,
              0.017281780019402504,
              1.2206934690475464,
              1.0955398082733154,
              0.01815686747431755,
              0.01785479672253132,
              0.03478417173027992,
              0.9184812903404236,
              1.572862148284912,
              0.7331628203392029,
              0.0465637631714344,
              0.6297574043273926,
              0.09811332076787949,
              0.01116716768592596,
              1.5792090892791748,
              0.38798901438713074,
              0.04957647621631622,
              0.01992660202085972,
              0.05843404680490494,
              0.004388938657939434,
              0.03892272338271141,
              0.010440375655889511,
              0.001601009746082127,
              0.004732479341328144,
              0.6406623721122742,
              0.5419203042984009,
              0.29596859216690063,
              9.05449104309082,
              0.27101460099220276,
              1.5104501247406006,
              0.029307130724191666,
              0.015437163412570953,
              1.2437081336975098,
              1.8386118412017822,
              1.6253098249435425,
              0.3509577512741089,
              0.0327141098678112,
              0.0615178719162941,
              0.03591759875416756,
              0.643612265586853,
              2.098417282104492,
              0.26738443970680237,
              0.11389615386724472,
              2.7116646766662598,
              0.687693178653717,
              0.5618981719017029,
              0.012411263771355152,
              0.00812049675732851,
              0.020271621644496918,
              0.00651450315490365,
              0.06100871041417122,
              0.15099835395812988,
              0.09459132701158524,
              0.2844947576522827,
              0.30277326703071594,
              0.008552944287657738,
              0.001983700320124626,
              0.006830083206295967,
              0.0474667027592659,
              1.740710973739624,
              0.4854883551597595,
              1.9961540699005127,
              1.4660340547561646,
              0.5130776166915894,
              5.532839298248291,
              7.031619071960449,
              7.962206840515137,
              0.7050954103469849,
              0.02929636463522911,
              0.00456837797537446,
              0.5282288789749146,
              0.3820890188217163,
              0.05044836178421974,
              0.002377542434260249,
              1.6788654327392578,
              0.47631216049194336,
              3.2770724296569824,
              1.4253441095352173,
              3.9487555027008057,
              0.9937187433242798,
              1.1734027862548828,
              0.8599129915237427,
              0.01017797365784645,
              0.0010193157941102982,
              0.13257113099098206,
              0.1615675538778305,
              0.011217622086405754,
              0.006486670114099979,
              0.01831243373453617,
              0.0004980515805073082,
              0.006735833827406168,
              9.107453346252441,
              0.25694912672042847,
              2.387226104736328,
              0.8998934626579285,
              0.00038842763751745224,
              0.009435346350073814,
              1.305315375328064,
              7.3980865478515625,
              0.005131526384502649,
              0.13160039484500885,
              0.028242671862244606,
              0.0007030876004137099,
              0.00707620196044445,
              0.10835885256528854,
              0.10640285909175873,
              0.010141038335859776,
              0.9094316363334656,
              0.027385415509343147,
              0.09689269959926605,
              6.692057132720947,
              0.2717764973640442,
              6.791062355041504,
              2.0940048694610596,
              0.21381087601184845,
              0.04802927374839783,
              2.421391248703003,
              9.32273006439209,
              0.03999534994363785,
              0.3151054084300995,
              3.056155204772949,
              0.09402961283922195,
              0.008975280448794365,
              0.0009039129945449531,
              0.8397841453552246,
              5.025770664215088,
              4.615621566772461,
              2.6083695888519287,
              0.20670680701732635,
              0.5682010650634766,
              0.874515175819397,
              2.0641939640045166,
              0.7174361348152161,
              0.07829786837100983,
              0.022562231868505478,
              5.1692280769348145,
              6.993217945098877,
              7.933887004852295,
              0.03295024856925011,
              0.34944555163383484,
              5.755311012268066,
              0.8602891564369202,
              0.00837433710694313,
              0.015615100041031837,
              0.0005066303419880569,
              0.7579731345176697,
              0.03532584756612778,
              0.011098437942564487,
              0.5613908767700195,
              1.117959976196289,
              0.6880788207054138,
              0.01624138467013836,
              1.5184569358825684,
              0.326661616563797,
              3.231532573699951,
              0.15875832736492157,
              0.9366632699966431,
              0.11091562360525131,
              0.0025527053512632847,
              0.007695906795561314,
              1.7117245197296143,
              0.005680963862687349,
              1.4774706363677979,
              0.6168534755706787,
              0.0028924793004989624,
              0.1830269694328308,
              0.04781655594706535,
              0.11460214108228683,
              0.0624229796230793,
              0.009019820019602776,
              0.010331599973142147,
              0.019788116216659546,
              0.0012419852428138256,
              6.875721454620361,
              1.9230016469955444,
              7.331838607788086,
              0.4252099096775055,
              0.36364924907684326,
              0.00404349947348237,
              0.9027749300003052,
              0.016164438799023628,
              5.856738567352295,
              0.49911415576934814,
              0.04592283070087433,
              0.12887001037597656,
              0.029800288379192352,
              0.0853920578956604,
              0.0028740549460053444,
              1.558236837387085,
              1.6369540691375732,
              0.34120723605155945,
              0.04081692174077034,
              0.20421874523162842,
              0.004791919607669115,
              1.7696548700332642,
              0.012758769094944,
              0.1438761204481125,
              0.01981125771999359,
              0.002046397188678384,
              0.0268936175853014,
              0.04710603505373001,
              0.16385935246944427,
              0.0318477489054203,
              0.0007459718617610633,
              1.3133518695831299,
              0.012470838613808155,
              0.01126866228878498,
              0.007856184616684914,
              0.004528267774730921,
              0.02784682996571064,
              1.1127396821975708,
              1.05778968334198,
              0.0664849653840065,
              0.002535464009270072,
              0.27774834632873535,
              3.008345127105713,
              0.6614764332771301,
              0.1744779348373413,
              0.0016221948899328709,
              0.022806724533438683,
              0.0012060280423611403,
              0.31103208661079407,
              0.0012657972984015942,
              0.022955160588026047,
              0.04872199520468712,
              0.004547373857349157,
              4.093289375305176,
              4.757928848266602,
              0.36956414580345154,
              7.174106597900391,
              3.126091480255127,
              6.393560886383057,
              1.0177867412567139,
              3.441581964492798,
              0.5863550901412964,
              5.9294257164001465,
              1.2209279537200928,
              1.7995719909667969,
              0.01860067807137966,
              0.0006587718962691724,
              0.23936496675014496,
              0.022070271894335747,
              5.0108771324157715,
              1.5817923545837402,
              2.2744064331054688,
              0.8287713527679443,
              0.6823278665542603,
              0.22025470435619354,
              0.4040948450565338,
              0.07114921510219574,
              0.03844849765300751,
              2.651553153991699,
              0.10412641614675522,
              0.14505580067634583,
              0.15453027188777924,
              0.00758316507562995,
              3.6314706802368164,
              5.673055171966553,
              0.9121759533882141,
              0.07465813308954239,
              1.713028907775879,
              0.019922861829400063,
              0.07012084126472473,
              0.024253802374005318,
              0.7852253317832947,
              0.35085225105285645,
              0.04621073976159096,
              1.8403822183609009,
              0.0003196682082489133,
              1.4305012882687151e-05,
              0.12860232591629028,
              0.21269303560256958,
              0.006926809437572956,
              0.40159493684768677,
              6.092898368835449,
              0.09558846056461334,
              1.3061774969100952,
              0.00018249277491122484,
              0.0003212174342479557,
              0.07394185662269592,
              0.0008317348547279835,
              0.027401423081755638,
              0.0007528808200731874,
              0.011328186839818954,
              0.004341937601566315,
              0.5559431314468384,
              0.010279921814799309,
              3.2598063945770264,
              0.1835893839597702,
              0.99027419090271,
              0.0014718662714585662,
              0.05761602520942688,
              0.7520503401756287,
              0.023628095164895058,
              0.15128891170024872,
              0.7915010452270508,
              0.0005570290377363563,
              0.9696894288063049,
              0.23200207948684692,
              0.5416104793548584,
              0.032183025032281876,
              0.0011039836099371314,
              8.797258487902582e-05,
              0.0016177912475541234,
              0.004669001325964928,
              0.005396442022174597,
              0.00051771110156551,
              0.0007556205382570624,
              0.007634391542524099,
              0.0009015309624373913,
              2.943364381790161,
              0.017758769914507866,
              0.028518103063106537,
              0.00030214988510124385,
              0.020879119634628296,
              0.0028278145473450422,
              0.0009950215462595224,
              0.006161503493785858,
              0.2904362380504608,
              0.0004533693427219987,
              0.5688651204109192,
              0.0022555168252438307,
              1.6926774978637695,
              3.541337728500366,
              1.2610138654708862,
              0.380769819021225,
              0.976841926574707,
              0.21846112608909607,
              0.000979539705440402,
              0.0015281677478924394,
              0.01708257757127285,
              0.11762820929288864,
              0.21288354694843292,
              0.010532978922128677,
              0.15096217393875122,
              0.0010644490830600262,
              4.275059700012207,
              0.11841512471437454,
              0.6050955653190613,
              6.055124282836914,
              0.05341324955224991,
              1.7589964866638184,
              0.016042560338974,
              0.0028067738749086857,
              0.7729501724243164,
              3.0454978942871094,
              0.005466512870043516,
              1.4402564764022827,
              0.04472146928310394,
              7.281580924987793,
              2.0297634601593018,
              0.0025128712877631187,
              0.01568751037120819,
              0.0010367024224251509,
              0.13267658650875092,
              0.0012677022023126483,
              1.0340900421142578,
              0.0023247378412634134,
              0.2905362844467163,
              0.002643188228830695,
              2.5466411113739014,
              0.10955669730901718,
              2.145662784576416,
              0.03284147381782532,
              0.009272251278162003,
              0.3050839602947235,
              0.01266860868781805,
              0.058449339121580124,
              1.4129787683486938,
              0.005802453961223364,
              0.00369437737390399,
              0.013313560746610165,
              0.0009005781612358987,
              1.070143222808838,
              2.5214009284973145,
              9.235360145568848,
              0.494171142578125,
              4.442967891693115,
              0.4019629657268524,
              1.9771836996078491,
              0.32261091470718384,
              0.007637467235326767,
              0.9558711051940918,
              0.22663846611976624,
              7.102785587310791,
              6.71352481842041,
              0.13189737498760223,
              0.01814129762351513,
              0.8429116010665894,
              1.8234776258468628,
              3.492260217666626,
              0.790324330329895,
              0.2718084454536438,
              0.23802198469638824,
              0.003978790249675512,
              0.16575147211551666,
              11.833624839782715,
              0.16591618955135345,
              0.17986206710338593,
              0.0055458261631429195,
              0.3585289418697357,
              0.0036454431246966124,
              2.133877992630005,
              0.31397566199302673,
              1.0533976554870605,
              0.704717218875885,
              3.654963254928589,
              1.120943307876587,
              7.829639911651611,
              5.781947135925293,
              0.042878516018390656,
              6.120713233947754,
              5.188594818115234,
              0.07457316666841507,
              3.632713794708252,
              6.4895734786987305,
              0.4927110970020294,
              1.2006816864013672,
              2.562880039215088,
              0.38893187046051025,
              1.137752890586853,
              0.0018385426374152303,
              0.3005598783493042,
              2.089118003845215,
              0.2551044225692749,
              0.07338226586580276,
              0.047065094113349915,
              0.02605692483484745,
              0.018409812822937965,
              0.9033674001693726,
              0.2610416114330292,
              0.017412647604942322,
              11.486262321472168,
              9.762709617614746,
              1.8505489826202393,
              11.85927963256836,
              1.3508386611938477,
              1.4082854986190796,
              0.7373759746551514,
              0.00895732268691063,
              0.32203832268714905,
              0.0003631647559814155,
              0.014993708580732346,
              0.27039939165115356,
              0.004002774599939585,
              0.11479420959949493,
              0.00916334893554449,
              5.015369415283203,
              1.524304986000061,
              1.3896381855010986,
              8.086215019226074,
              1.8132929801940918,
              0.019564513117074966,
              7.052774906158447,
              1.1906909942626953,
              7.081691265106201,
              0.6857237219810486,
              0.003565622726455331,
              0.01653151400387287,
              0.7350987195968628,
              0.9738313555717468,
              0.336582750082016,
              0.3849683403968811,
              5.7063703536987305,
              0.40877997875213623,
              0.002282991772517562,
              0.0037664680276066065,
              0.1660735160112381,
              0.1385461837053299,
              0.002397878561168909,
              0.0002044230350293219,
              0.06832226365804672,
              0.003880472155287862,
              0.32116663455963135,
              0.13392199575901031,
              0.0028368488419800997,
              0.002273952355608344,
              0.0007420408655889332,
              0.00018976318824570626,
              0.0092361094430089,
              0.0003721021639648825,
              0.0013830630341544747,
              0.0005029367166571319,
              0.25371885299682617,
              0.004007167648524046,
              15.061026573181152,
              0.029593611136078835,
              8.691543579101562,
              1.1829710006713867,
              0.012860337272286415,
              0.9301536083221436,
              0.0015298341168090701,
              0.003188884584233165,
              0.3979724645614624,
              0.8979056477546692,
              0.0021371636539697647,
              0.17863911390304565,
              0.06433705240488052,
              0.0007384672062471509,
              0.04878046736121178,
              1.7046782886609435e-05,
              0.004256949760019779,
              0.00524691678583622,
              0.3824421167373657,
              0.0015591143164783716,
              0.0004892344586551189,
              0.002014156896620989,
              0.640900731086731,
              0.005283796228468418,
              0.008050612173974514,
              0.0012531769461929798,
              0.00031799983116798103,
              0.034635063260793686,
              0.06391610950231552,
              0.002556034829467535,
              0.5099653601646423,
              0.0003067976504098624,
              0.001257344032637775,
              0.00035446559195406735,
              0.04389968141913414,
              0.003462512046098709,
              0.9078320264816284,
              0.006725413724780083,
              0.5767203569412231,
              0.9164893627166748,
              0.2063876986503601,
              0.2067936807870865,
              0.00032145579461939633,
              0.0001102625101339072,
              0.0005729940603487194,
              0.0013013951247557998,
              0.0028028511442244053,
              0.23761138319969177,
              2.027841091156006,
              0.6806821227073669,
              7.578509330749512,
              7.880727291107178,
              0.5352344512939453,
              2.706234931945801,
              1.3803693056106567,
              0.17478016018867493,
              0.011698336340487003,
              0.04247711971402168,
              0.048449575901031494,
              0.0017714299028739333,
              5.134828567504883,
              0.10553178191184998,
              0.7957468628883362,
              0.7811497449874878,
              0.25955694913864136,
              0.0727454423904419,
              7.434291362762451,
              0.18680551648139954,
              4.9988017082214355,
              9.935079574584961,
              0.344137966632843,
              6.096242427825928,
              1.5712672472000122,
              0.0013869914691895247,
              0.0038791659753769636,
              1.675551176071167,
              0.15461668372154236,
              0.0014488924061879516,
              0.0003045333724003285,
              0.06425075232982635,
              0.0031708222813904285,
              0.2664250433444977,
              0.003941031638532877,
              0.01026446558535099,
              0.00045813556062057614,
              0.001166854752227664,
              0.01743701472878456,
              0.0002584123576525599,
              0.003457641461864114,
              0.020596904680132866,
              0.05025169253349304,
              0.000724887300748378,
              0.006422829814255238,
              0.12891989946365356,
              0.017127107828855515,
              0.011645196937024593,
              0.002873579505831003,
              15.025128364562988,
              0.1084240972995758,
              0.010265881195664406,
              0.000720956246368587,
              0.0008454324561171234,
              0.0015432840446010232,
              0.025814039632678032,
              0.012821970507502556,
              8.160237312316895,
              0.0011503038695082068,
              5.788867950439453,
              0.0033094894606620073,
              0.005120615474879742,
              0.0006690170848742127,
              0.22546011209487915,
              3.218599158572033e-05,
              7.092700980138034e-05,
              0.003561346558853984,
              0.0005243833875283599,
              0.00011646069469861686,
              0.03904024139046669,
              0.0005947966128587723,
              0.01298319362103939,
              0.0010355116100981832,
              0.03540213778614998,
              0.0007128558354452252,
              0.0030607779044657946,
              8.594620157964528e-05,
              4.792098479811102e-05,
              0.0007744412287138402,
              0.00034791138023138046,
              0.1828223466873169,
              0.32692229747772217,
              0.054980698972940445,
              0.16217632591724396,
              0.0006163604557514191,
              0.0006735440110787749,
              0.009371339343488216,
              0.0011018402874469757,
              3.909505844116211,
              2.4928343296051025,
              0.22566017508506775,
              7.757230758666992,
              0.904852032661438,
              0.005460940767079592,
              0.0017498909728601575,
              0.45829087495803833,
              0.06347864866256714,
              0.16344286501407623,
              1.77726149559021,
              0.022963082417845726,
              0.148712620139122,
              0.3148076832294464,
              3.7633533477783203,
              6.183507442474365,
              0.1742970049381256,
              2.237128734588623,
              4.538725852966309,
              0.5947341918945312,
              0.1507757008075714,
              1.0251219272613525,
              4.231703758239746,
              9.741634368896484,
              8.304247856140137,
              0.8171319961547852,
              0.012018527835607529,
              0.0008760428754612803,
              0.7538920640945435,
              0.03375422954559326,
              0.5605875253677368,
              0.009123304858803749,
              0.03263138607144356,
              0.7920892834663391,
              0.006535702850669622,
              1.9576035737991333,
              0.019991805776953697,
              0.0397881343960762,
              4.868468284606934,
              0.9831074476242065,
              0.018613314256072044,
              0.02297857776284218,
              0.4092073142528534,
              0.024444719776511192,
              0.12283331900835037,
              0.10178878158330917,
              2.1762311458587646,
              0.25880205631256104,
              0.0014693664852529764,
              0.0217581819742918,
              0.001671466277912259,
              11.013932228088379,
              1.9765872955322266,
              0.5634450316429138,
              0.6884610056877136,
              0.07229187339544296,
              0.00041631137719377875,
              8.758684158325195,
              0.3169921636581421,
              0.12192768603563309,
              0.12404545396566391,
              0.19380038976669312,
              0.0025598397478461266,
              2.3704304695129395,
              0.01949787698686123,
              0.10137071460485458,
              0.00693225534632802,
              1.346453070640564,
              0.00648216949775815,
              0.5920891165733337,
              0.30761125683784485,
              3.8414738178253174,
              0.21796314418315887,
              0.02804424799978733,
              0.06732608377933502,
              0.06329985707998276,
              0.06431547552347183,
              0.001490316353738308,
              0.0917031466960907,
              0.06218495965003967,
              1.2823399305343628,
              0.1261129230260849,
              0.08208420127630234,
              0.003927020356059074,
              0.0035121680703014135,
              0.1446543037891388,
              0.0015817285748198628,
              0.06966341286897659,
              0.029600324109196663,
              1.7696847915649414,
              4.772102355957031,
              0.2724061608314514,
              0.008632130920886993,
              2.9936840534210205,
              0.04015099257230759,
              0.07470260560512543,
              0.0021634523291140795,
              0.0033410938922315836,
              0.28520333766937256,
              0.002842554822564125,
              2.9906723499298096,
              3.731140613555908,
              0.29857349395751953,
              0.07795587182044983,
              0.48902422189712524,
              0.024230647832155228,
              0.5852993130683899,
              0.004078523721545935,
              1.5437309741973877,
              0.14815141260623932,
              7.3670783042907715,
              0.48644477128982544,
              8.001540184020996,
              1.1641314029693604,
              0.02092861942946911,
              5.693403244018555,
              0.05125663802027702,
              0.3546053469181061,
              4.964481353759766,
              5.20483922958374,
              1.2226676940917969,
              0.7787058353424072,
              0.0030269066337496042,
              2.4134511947631836,
              0.039405547082424164,
              6.020589828491211,
              10.137667655944824,
              0.17680279910564423,
              0.830873966217041,
              1.3913600444793701,
              0.26253873109817505,
              0.30726632475852966,
              0.001292942208237946,
              1.5529899597167969,
              2.3181140422821045,
              0.03447948023676872,
              5.88485050201416,
              1.140701174736023,
              6.560232162475586,
              6.520287990570068,
              0.1597450226545334,
              1.6155089139938354,
              0.16646640002727509,
              0.049461666494607925,
              0.0030711174476891756,
              7.8635573387146,
              0.4338708519935608,
              0.0013550871517509222,
              0.005397627595812082,
              0.0007321537705138326,
              2.220344066619873,
              1.4188683032989502,
              2.077694892883301,
              0.0030409307219088078,
              0.023482901975512505,
              1.6139737367630005,
              0.1224723681807518,
              0.034719351679086685,
              0.006644062697887421,
              4.338037967681885,
              3.50416898727417,
              8.57374095916748,
              0.13536079227924347,
              0.0018299752846360207,
              0.010605285875499249,
              0.0012903229799121618,
              4.373175621032715,
              0.2017698884010315,
              2.081063747406006,
              0.24561429023742676,
              2.509530544281006,
              0.22902099788188934,
              4.659517288208008,
              2.485172986984253,
              3.357473850250244,
              0.4271472096443176,
              0.06526362150907516,
              0.5774276852607727,
              0.08758144080638885,
              0.0741177648305893,
              0.36058688163757324,
              1.8344959020614624,
              0.19233858585357666,
              2.233283758163452,
              0.1622292399406433,
              0.5341155529022217,
              0.020163560286164284,
              0.13763733208179474,
              0.0009479796281084418,
              0.367675244808197,
              0.18558627367019653,
              0.1630728840827942,
              0.0003262225945945829,
              0.008776542730629444,
              0.7169497609138489,
              0.004523165058344603,
              4.61142110824585,
              0.5674894452095032,
              0.0550866425037384,
              0.8033467531204224,
              0.13231906294822693,
              0.0038887844420969486,
              1.7124874591827393,
              0.0020898187067359686,
              0.9397075772285461,
              0.0028247239533811808,
              0.004612757824361324,
              0.005728020798414946,
              0.47983306646347046,
              1.7003567218780518,
              5.742242813110352,
              2.175626516342163,
              0.25671520829200745,
              2.1866023540496826,
              0.018553519621491432,
              2.289531707763672,
              6.347573280334473,
              0.42354899644851685,
              0.31790071725845337,
              0.02005525305867195,
              0.12263203412294388,
              0.04115217924118042,
              0.2132263034582138,
              1.1361019611358643,
              2.642930030822754,
              6.036240577697754,
              0.02378525771200657,
              0.11229262501001358,
              0.05106533318758011,
              0.3921520411968231,
              0.00033802041434682906,
              0.0022695516236126423,
              0.15350395441055298,
              0.10364827513694763,
              0.364859938621521,
              1.2477688789367676,
              0.004083035048097372,
              8.039987564086914,
              0.3342653512954712,
              0.0012759171659126878,
              0.010327942669391632,
              0.0014399646315723658,
              3.5133140087127686,
              2.2340569496154785,
              3.0631790161132812,
              0.000942858459893614,
              2.6357247829437256,
              2.8013839721679688,
              0.8668656349182129,
              0.010071411728858948,
              1.0168161392211914,
              0.978556752204895,
              1.230711817741394,
              0.08852546662092209,
              0.8543406128883362
            ],
            "kl": [],
            "kl_count": 0,
            "ok": true,
            "seconds": 1.6894906880000065,
            "gpu_observation": {
              "scope": "whole_device",
              "sample_interval_seconds": 0.1,
              "samples": 16,
              "start_used_bytes": 30730846208,
              "peak_used_bytes": 33393115136,
              "end_used_bytes": 33393115136,
              "max_temperature_c": 38,
              "max_power_w": 77.746
            }
          }
        },
        {
          "id": "validation-001",
          "max_abs_token_nll_change": 0.0,
          "result": {
            "id": "validation-001",
            "nll_sum": 2708.6642821885343,
            "scored_tokens": 1023,
            "token_nll": [
              9.52013874053955,
              12.002548217773438,
              2.2489867210388184,
              7.736156940460205,
              6.982629776000977,
              7.083405494689941,
              2.1704444885253906,
              5.45363712310791,
              0.6513281464576721,
              0.49308276176452637,
              0.5399232506752014,
              0.26831501722335815,
              4.537866592407227,
              5.843966007232666,
              1.4133515357971191,
              0.48656949400901794,
              0.48293280601501465,
              8.686925888061523,
              2.7491402626037598,
              9.830156326293945,
              1.415894627571106,
              0.6228544116020203,
              0.00804316159337759,
              3.3856637477874756,
              0.9824801683425903,
              4.771305084228516,
              1.818648099899292,
              3.347621202468872,
              3.4381213188171387,
              1.0415493249893188,
              4.06350564956665,
              2.4780726432800293,
              7.780920505523682,
              1.0956788063049316,
              6.527156829833984,
              4.360079765319824,
              3.169369697570801,
              4.763039588928223,
              2.159048557281494,
              4.061174392700195,
              0.10069207847118378,
              0.8176100850105286,
              0.6291301250457764,
              1.9400837421417236,
              0.8991982936859131,
              3.692934989929199,
              0.35381919145584106,
              2.080996036529541,
              0.008905455470085144,
              0.6580124497413635,
              2.7285499572753906,
              7.402239799499512,
              4.475652694702148,
              1.0794143676757812,
              0.004874963779002428,
              1.2453209161758423,
              7.106649875640869,
              2.986445903778076,
              2.411364793777466,
              3.4426541328430176,
              9.921222686767578,
              1.1427440643310547,
              0.9658037424087524,
              0.24190396070480347,
              0.03176495060324669,
              4.765766143798828,
              0.302436888217926,
              0.04982035234570503,
              0.05247845500707626,
              0.016262026503682137,
              1.9010759592056274,
              1.2622658014297485,
              4.2578325271606445,
              0.39999738335609436,
              0.04387321323156357,
              0.0128423310816288,
              2.3639163970947266,
              0.19565477967262268,
              2.530421495437622,
              1.1852247714996338,
              9.045199394226074,
              2.6389167308807373,
              2.99233341217041,
              2.3178443908691406,
              4.4180908203125,
              2.601801872253418,
              6.5359673500061035,
              1.722730040550232,
              0.5832419991493225,
              4.109786033630371,
              1.567089319229126,
              0.24157801270484924,
              0.62485271692276,
              2.3370957374572754,
              6.058615207672119,
              2.318607807159424,
              0.035741399973630905,
              0.01756657287478447,
              1.1621527671813965,
              1.1670506000518799,
              2.241014003753662,
              1.362016201019287,
              0.6177926659584045,
              0.05230933427810669,
              1.2536154985427856,
              2.481278657913208,
              0.06438980996608734,
              2.299549102783203,
              1.5497442483901978,
              1.7808101177215576,
              3.0146775245666504,
              3.772451877593994,
              1.4937748908996582,
              0.1082671731710434,
              5.252569198608398,
              0.21217048168182373,
              0.0007173825288191438,
              0.008874026127159595,
              0.37461069226264954,
              0.03351124748587608,
              2.449326515197754,
              0.6379566788673401,
              0.29393377900123596,
              0.9443280696868896,
              2.4744648933410645,
              1.1931442022323608,
              2.1352148056030273,
              1.0994985103607178,
              0.006576442625373602,
              2.4022374153137207,
              1.7561386823654175,
              0.04525288939476013,
              8.73627758026123,
              1.6380332708358765,
              2.191119909286499,
              10.151932716369629,
              5.58559513092041,
              2.583759307861328,
              0.9925829172134399,
              5.943935871124268,
              6.607207298278809,
              2.514878749847412,
              3.8420562744140625,
              2.4225268363952637,
              3.752835750579834,
              4.020107746124268,
              1.1709790229797363,
              3.625967025756836,
              0.21045805513858795,
              2.719209909439087,
              4.507439613342285,
              6.339090347290039,
              0.09589870274066925,
              0.0006940339808352292,
              0.00010215714428341016,
              0.09014923125505447,
              0.019719859585165977,
              1.0698810815811157,
              0.20958366990089417,
              1.6655155420303345,
              0.256828248500824,
              0.5704367160797119,
              3.806450366973877,
              0.5062576532363892,
              0.23394525051116943,
              0.04056317359209061,
              0.04961981251835823,
              1.046252965927124,
              0.09029911458492279,
              0.1302337348461151,
              4.780104160308838,
              0.22014681994915009,
              0.6323316097259521,
              0.771594762802124,
              0.015376473776996136,
              9.026691436767578,
              5.1473565101623535,
              1.3199158906936646,
              1.7405927181243896,
              0.9141393303871155,
              4.431920528411865,
              0.027568893507122993,
              0.14266568422317505,
              4.015029430389404,
              3.83156681060791,
              0.09014759957790375,
              0.02191493846476078,
              5.682604789733887,
              0.3811469078063965,
              1.115633249282837,
              1.4747169017791748,
              0.001755721983499825,
              0.3703579604625702,
              1.4685672521591187,
              0.09185029566287994,
              1.7258481979370117,
              7.845141410827637,
              2.809051036834717,
              9.446828842163086,
              4.171982765197754,
              2.538567543029785,
              6.074326515197754,
              4.786511421203613,
              2.44923734664917,
              3.678257465362549,
              2.472917079925537,
              8.326075553894043,
              2.1317172050476074,
              4.014934062957764,
              0.4702262282371521,
              4.5621232986450195,
              6.087001800537109,
              6.460465908050537,
              0.7479300498962402,
              4.620270252227783,
              5.405426025390625,
              0.5368247628211975,
              7.860622882843018,
              2.6945300102233887,
              2.4484665393829346,
              4.52205753326416,
              0.4001796543598175,
              0.18742570281028748,
              0.015051954425871372,
              2.708491086959839,
              2.2739739418029785,
              6.1473188400268555,
              0.8747371435165405,
              0.04701732471585274,
              0.0035943682305514812,
              5.706228256225586,
              3.885612964630127,
              0.005488445982336998,
              0.02589663304388523,
              4.088173866271973,
              4.145544528961182,
              0.4073341488838196,
              0.03224685788154602,
              1.4364643096923828,
              6.434864044189453,
              8.021523475646973,
              0.041275039315223694,
              0.3826826214790344,
              1.5614206790924072,
              0.3057628571987152,
              4.033501625061035,
              1.9354864358901978,
              2.1410155296325684,
              6.892793655395508,
              0.42614108324050903,
              7.289139747619629,
              6.081064224243164,
              6.611526966094971,
              3.300208330154419,
              0.1674354523420334,
              0.22724175453186035,
              6.141839504241943,
              0.30532142519950867,
              2.909240245819092,
              3.7135496139526367,
              0.0416128896176815,
              2.6620123386383057,
              6.482154846191406,
              2.7307634353637695,
              1.4799892902374268,
              0.00386265991255641,
              0.520313024520874,
              1.2054260969161987,
              0.06755954027175903,
              1.430454969406128,
              9.102563858032227,
              6.219606399536133,
              0.9666482210159302,
              6.733099460601807,
              4.165942668914795,
              4.390035629272461,
              0.6812820434570312,
              4.699453353881836,
              2.7545864582061768,
              0.3632604479789734,
              5.953596591949463,
              3.083885431289673,
              6.075586795806885,
              0.04576011374592781,
              5.4624247550964355,
              3.7884788513183594,
              2.235166549682617,
              1.3392868041992188,
              4.152890205383301,
              3.6835975646972656,
              0.6967176795005798,
              5.094457626342773,
              8.776686668395996,
              0.3436611294746399,
              0.21551911532878876,
              0.003585103200748563,
              6.546798229217529,
              3.3380300998687744,
              0.07188580185174942,
              2.8937230110168457,
              0.7332068085670471,
              2.54687762260437,
              3.7017414569854736,
              2.196678876876831,
              0.03479718416929245,
              1.755513310432434,
              6.002725601196289,
              2.1665854454040527,
              0.013308149762451649,
              0.0007167869480326772,
              0.0033199451863765717,
              0.15833714604377747,
              0.9343689680099487,
              4.008975505828857,
              3.011183977127075,
              0.06343680620193481,
              1.399986982345581,
              7.201993942260742,
              3.7578301429748535,
              5.920069694519043,
              2.7039589881896973,
              0.011523472145199776,
              0.005259842611849308,
              0.011257699690759182,
              8.002809524536133,
              0.4713149070739746,
              9.39640998840332,
              2.0291097164154053,
              0.016703981906175613,
              5.01519775390625,
              0.21186411380767822,
              0.03029482252895832,
              0.4888567328453064,
              0.003639266826212406,
              12.012791633605957,
              3.4073851108551025,
              1.3049319982528687,
              2.0471951961517334,
              5.640341758728027,
              5.181522369384766,
              1.0265579223632812,
              5.712615013122559,
              0.02745083160698414,
              0.00933496467769146,
              0.3627079427242279,
              4.041610240936279,
              2.598806381225586,
              3.533130407333374,
              6.305349826812744,
              3.697490930557251,
              4.097785949707031,
              4.896533489227295,
              0.10606230795383453,
              3.846529483795166,
              6.385831832885742,
              8.872353553771973,
              2.8422493934631348,
              5.199817657470703,
              4.3152971267700195,
              1.1819381713867188,
              0.010937736369669437,
              0.22587409615516663,
              0.001970613142475486,
              2.647271156311035,
              0.07964608073234558,
              4.988974571228027,
              1.4239040613174438,
              6.392162322998047,
              0.885921061038971,
              1.6195621490478516,
              2.7298929691314697,
              0.056169018149375916,
              0.7875789403915405,
              0.023407327011227608,
              5.160245895385742,
              4.225543975830078,
              5.985058307647705,
              5.418036937713623,
              0.0793180912733078,
              1.556401014328003,
              3.3264482021331787,
              7.322381973266602,
              2.9751129150390625,
              4.969808101654053,
              0.23901031911373138,
              1.301679253578186,
              0.004251252394169569,
              2.4318398573086597e-05,
              0.004957051947712898,
              2.4136505126953125,
              1.849637746810913,
              0.11040934175252914,
              0.15333062410354614,
              2.0718252658843994,
              5.133758068084717,
              0.0014173472300171852,
              4.864749431610107,
              3.3845767974853516,
              8.729533195495605,
              0.6695420742034912,
              6.517980098724365,
              2.640540599822998,
              3.370121955871582,
              0.900827944278717,
              10.78148078918457,
              1.23940908908844,
              0.6421162486076355,
              8.539202690124512,
              7.126778602600098,
              0.02781854197382927,
              0.12017833441495895,
              3.43646502494812,
              0.9788106679916382,
              5.2755961418151855,
              1.2839224338531494,
              0.33191701769828796,
              1.9831781387329102,
              0.10616854578256607,
              1.5108072757720947,
              1.3412882089614868,
              1.564582109451294,
              0.19084884226322174,
              0.26265910267829895,
              0.0015185264637693763,
              5.23579740524292,
              1.8870505094528198,
              0.02434990555047989,
              1.5293018817901611,
              6.365276336669922,
              3.602062702178955,
              4.072422027587891,
              1.8939564228057861,
              2.483518362045288,
              0.9312278032302856,
              2.547039747238159,
              2.254912853240967,
              0.01946946792304516,
              0.0046170298010110855,
              0.797711968421936,
              9.116186141967773,
              1.9671567678451538,
              2.3043582439422607,
              4.265352725982666,
              1.1981608867645264,
              0.0033617669250816107,
              4.075191020965576,
              6.724794864654541,
              5.9395856857299805,
              1.0080236196517944,
              0.329090416431427,
              0.909895122051239,
              0.014253122732043266,
              1.4007568359375,
              1.2435028553009033,
              0.04503490775823593,
              1.347987174987793,
              8.34876823425293,
              5.598801136016846,
              0.4839744567871094,
              6.667313575744629,
              2.858527898788452,
              6.0194854736328125,
              0.5928025245666504,
              2.49735164642334,
              5.852606296539307,
              1.4082305431365967,
              1.5884265899658203,
              2.2071399688720703,
              0.12587876617908478,
              0.05913679301738739,
              3.3802742958068848,
              1.94303560256958,
              0.04355975240468979,
              4.200966835021973,
              2.070208787918091,
              0.2658049166202545,
              4.885797500610352,
              1.5254942178726196,
              0.08532758802175522,
              7.575620651245117,
              2.584751605987549,
              4.6125078201293945,
              1.2440402507781982,
              4.306641578674316,
              6.79020881652832,
              3.3119869232177734,
              1.455621361732483,
              4.180821418762207,
              1.880200982093811,
              0.9293673038482666,
              4.681321620941162,
              2.633932113647461,
              8.6801176071167,
              3.2953503131866455,
              3.534891128540039,
              8.849013328552246,
              3.227917194366455,
              8.627224922180176,
              1.8106496334075928,
              3.339733362197876,
              0.1411752700805664,
              2.240239381790161,
              9.36635971069336,
              0.39735886454582214,
              0.0013435394503176212,
              0.003944593947380781,
              1.2565886974334717,
              3.667970657348633,
              2.355546236038208,
              0.33492982387542725,
              0.006448769476264715,
              2.6963937282562256,
              2.8459632396698,
              0.36682966351509094,
              0.021082470193505287,
              0.5462791323661804,
              3.415285110473633,
              0.0002479245886206627,
              0.00032395837479270995,
              1.6813945770263672,
              0.0017300175968557596,
              2.6758229732513428,
              4.184305667877197,
              10.729701042175293,
              0.4366479814052582,
              4.48606014251709,
              2.35199236869812,
              5.4959025382995605,
              8.303703308105469,
              2.243964195251465,
              0.9387340545654297,
              2.0472323894500732,
              6.15701961517334,
              0.8106436729431152,
              0.015663335099816322,
              2.071047782897949,
              4.933618068695068,
              2.675654411315918,
              9.778838157653809,
              0.20248231291770935,
              3.822169780731201,
              4.957231521606445,
              0.3530864417552948,
              0.0017099058022722602,
              2.0101895332336426,
              0.07595711946487427,
              3.4656996726989746,
              1.4002364873886108,
              7.853762626647949,
              0.24471254646778107,
              7.311191082000732,
              1.3148590326309204,
              6.424290657043457,
              3.621105670928955,
              10.533852577209473,
              11.050724029541016,
              0.431475430727005,
              8.332113265991211,
              7.237595558166504,
              3.1808056831359863,
              0.7082105278968811,
              0.7707714438438416,
              0.007443787530064583,
              5.135791778564453,
              9.666401863098145,
              2.606980800628662,
              4.553239822387695,
              0.044174257665872574,
              6.629244804382324,
              0.01932811364531517,
              3.3037967681884766,
              2.5865366458892822,
              2.6291961669921875,
              9.131392478942871,
              5.3236823081970215,
              1.7625367641448975,
              1.7612687349319458,
              3.423616409301758,
              5.459888458251953,
              0.1729235202074051,
              10.438714027404785,
              4.035504341125488,
              0.026906035840511322,
              0.003479024861007929,
              9.116545677185059,
              0.21822510659694672,
              7.562967300415039,
              0.014614555984735489,
              1.701676607131958,
              1.4903544187545776,
              9.701549530029297,
              5.234119415283203,
              0.0800316333770752,
              0.0036674162838608027,
              0.32970210909843445,
              0.16542619466781616,
              1.5382198095321655,
              2.888974905014038,
              3.7731289863586426,
              2.411900520324707,
              2.350750684738159,
              3.102579116821289,
              4.538564205169678,
              3.6543455123901367,
              5.9579572677612305,
              0.841665506362915,
              0.30645880103111267,
              1.5495028495788574,
              0.8527650833129883,
              0.775176465511322,
              4.83044958114624,
              4.15620231628418,
              3.8427681922912598,
              1.3022202253341675,
              0.002186290919780731,
              0.34646540880203247,
              8.748939514160156,
              4.390158653259277,
              1.5454524755477905,
              0.061212360858917236,
              4.8776631355285645,
              5.541364669799805,
              2.3010575771331787,
              1.9942538738250732,
              1.2197606563568115,
              4.999879837036133,
              9.67459774017334,
              4.175739765167236,
              0.8834035396575928,
              3.6236062049865723,
              1.0331720113754272,
              1.617417335510254,
              0.18709634244441986,
              0.051303863525390625,
              0.0035947246942669153,
              0.7667838931083679,
              2.165177583694458,
              3.2173190116882324,
              1.4046707153320312,
              4.654242515563965,
              3.0834295749664307,
              0.024394812062382698,
              3.847564935684204,
              1.0801135301589966,
              2.9893763065338135,
              0.24650366604328156,
              2.019948720932007,
              6.755216598510742,
              2.146714210510254,
              7.214040279388428,
              7.8522047996521,
              0.00949143711477518,
              1.3611578941345215,
              7.005184173583984,
              6.7560224533081055,
              1.2511132955551147,
              6.1092681884765625,
              0.5832107067108154,
              0.9563677310943604,
              4.556482315063477,
              1.4577008485794067,
              0.010033882223069668,
              5.412189960479736,
              1.1718896627426147,
              2.8535752296447754,
              0.7666105031967163,
              7.922016620635986,
              0.4437370300292969,
              0.4853051006793976,
              1.8929996490478516,
              5.172942161560059,
              1.217557430267334,
              4.61669921875,
              3.7285876274108887,
              3.132673740386963,
              4.729178428649902,
              0.16322274506092072,
              2.800417900085449,
              0.3483791947364807,
              0.528766393661499,
              0.0032086102291941643,
              10.044210433959961,
              9.167045593261719,
              7.302003860473633,
              2.3860082626342773,
              3.244396209716797,
              7.9691667556762695,
              3.7876617908477783,
              1.7936397790908813,
              0.00299232080578804,
              5.772668838500977,
              12.293994903564453,
              0.05114711448550224,
              3.735440731048584,
              4.050382137298584,
              3.0286755561828613,
              3.6071579456329346,
              0.7645921111106873,
              5.605196952819824,
              0.08095257729291916,
              1.7436370849609375,
              0.8199892044067383,
              0.20012736320495605,
              2.8581531047821045,
              5.08929967880249,
              9.155624389648438,
              3.307541847229004,
              5.861201286315918,
              5.962027549743652,
              2.1287550926208496,
              0.5144307613372803,
              2.4444282054901123,
              0.008435807190835476,
              1.048807978630066,
              1.3321068286895752,
              5.015321254730225,
              6.9317731857299805,
              8.245481491088867,
              3.149745464324951,
              0.006526583340018988,
              4.500169277191162,
              6.417437553405762,
              8.66666030883789,
              3.3015878200531006,
              1.6830711364746094,
              2.9286909103393555,
              8.99172592163086,
              0.4224042296409607,
              0.06270964443683624,
              0.5819514393806458,
              3.0353498458862305,
              3.7481110095977783,
              0.15250737965106964,
              1.6968241930007935,
              0.49913525581359863,
              0.6820584535598755,
              0.002999095479026437,
              6.936326503753662,
              0.24210704863071442,
              0.5040844678878784,
              0.41803231835365295,
              4.3634772300720215,
              0.5571714639663696,
              1.0179333686828613,
              0.12600195407867432,
              8.262211799621582,
              3.8760733604431152,
              2.3318140506744385,
              0.7588400840759277,
              0.3598584830760956,
              4.252270698547363,
              4.130010604858398,
              7.038558006286621,
              3.388709783554077,
              6.3310770988464355,
              8.106330871582031,
              0.11122968047857285,
              2.0583372116088867,
              0.6294359564781189,
              6.721256256103516,
              4.303809642791748,
              6.255277156829834,
              5.205979347229004,
              1.839691162109375,
              0.00540023623034358,
              2.1055078506469727,
              1.2659471035003662,
              0.47792041301727295,
              0.0022185256239026785,
              0.7580448985099792,
              4.191143989562988,
              0.02819330058991909,
              1.014024257659912,
              1.3445292711257935,
              0.027044234797358513,
              1.2161685228347778,
              0.44733965396881104,
              1.538764476776123,
              1.8816026449203491,
              0.9103418588638306,
              0.12783420085906982,
              0.000644237850792706,
              1.054574966430664,
              0.0048888432793319225,
              0.6401573419570923,
              0.5371537804603577,
              2.061683177947998,
              0.9314525127410889,
              4.157558441162109,
              0.2316526472568512,
              4.421121597290039,
              1.0113506317138672,
              0.7369930148124695,
              5.017963409423828,
              3.6184182167053223,
              0.015175716951489449,
              0.5572605729103088,
              0.014065656810998917,
              6.649529457092285,
              9.322870254516602,
              0.73996502161026,
              1.9200117588043213,
              0.4597889184951782,
              4.970315933227539,
              8.493839263916016,
              0.043285008519887924,
              0.0024915861431509256,
              0.0028178293723613024,
              0.0017830915749073029,
              0.6787890791893005,
              0.2686477601528168,
              6.296095848083496,
              0.11258047819137573,
              2.661877155303955,
              0.5130460858345032,
              1.0482263565063477,
              4.230615139007568,
              0.010605758056044579,
              2.126100540161133,
              7.92125129699707,
              0.07716592401266098,
              2.815237045288086,
              0.013848294503986835,
              2.36625337600708,
              2.677311897277832,
              0.6388340592384338,
              7.439090251922607,
              6.081787586212158,
              5.438658714294434,
              0.6017703413963318,
              4.865327835083008,
              0.33695346117019653,
              5.831650257110596,
              0.3667183816432953,
              0.004174566827714443,
              2.9191391468048096,
              0.15840990841388702,
              0.008652222342789173,
              0.8877695202827454,
              0.15600532293319702,
              1.324544906616211,
              2.636751651763916,
              0.03006337769329548,
              0.06593658775091171,
              0.0004462200158741325,
              0.03153650090098381,
              0.00023457636416424066,
              1.4341776371002197,
              1.7278242111206055,
              1.4976445436477661,
              1.7852144241333008,
              3.663688898086548,
              0.5727733373641968,
              10.008241653442383,
              0.3660668432712555,
              6.698319911956787,
              5.79608678817749,
              0.5687733292579651,
              5.901826858520508,
              1.2719409465789795,
              0.47378110885620117,
              3.4874439239501953,
              1.9948428869247437,
              4.514664649963379,
              1.4556742906570435,
              1.1260191202163696,
              0.2614597678184509,
              6.348115921020508,
              1.107658863067627,
              0.21057626605033875,
              4.469581604003906,
              0.7620528936386108,
              2.079153299331665,
              10.954011917114258,
              0.858432412147522,
              0.07731357216835022,
              2.9709668159484863,
              1.1911300420761108,
              1.7320449352264404,
              2.754031181335449,
              4.077364444732666,
              4.807592868804932,
              0.30890265107154846,
              2.7425730228424072,
              0.5119137167930603,
              5.150711536407471,
              2.3792307376861572,
              0.12903495132923126,
              0.0034893599804490805,
              1.4003878831863403,
              0.7525083422660828,
              0.0027019204571843147,
              0.9180143475532532,
              0.9478853940963745,
              2.246816396713257,
              3.6573193073272705,
              4.031844139099121,
              0.10876214504241943,
              0.11821845173835754,
              0.7130547165870667,
              1.147607684135437,
              0.37226980924606323,
              1.686330795288086,
              0.49717944860458374,
              0.4750874936580658,
              3.8958003520965576,
              5.7793402671813965,
              1.0843322277069092,
              2.8691487312316895,
              1.3456509113311768,
              0.2756218910217285,
              3.369493007659912,
              5.262314319610596,
              7.03236722946167,
              2.123485803604126,
              0.00014995403762441128,
              1.9857962131500244,
              0.00821745302528143,
              3.265122890472412,
              0.1462668627500534,
              0.007099401671439409,
              0.22979623079299927,
              1.8143384456634521,
              2.4211177825927734,
              2.415961742401123,
              8.10753059387207,
              1.218666434288025,
              2.2713751792907715,
              1.544224500656128,
              1.446012258529663,
              0.2787134647369385,
              1.5722310543060303,
              0.18628330528736115,
              2.936514139175415,
              0.09915175288915634,
              4.222903728485107,
              1.8265373706817627,
              4.7434773445129395,
              0.5559893846511841,
              0.051736149936914444,
              0.6992998123168945,
              0.42817407846450806,
              0.005474811885505915,
              0.0005536930402740836,
              0.013994187116622925,
              0.00807272456586361,
              2.0782694816589355,
              1.600785255432129,
              0.8166961073875427,
              1.622453212738037,
              1.3907418251037598,
              0.5150018334388733,
              0.16946382820606232,
              0.011929478496313095,
              7.237950801849365,
              8.364789009094238,
              2.4797284603118896,
              0.23961369693279266,
              0.0038330908864736557,
              2.5522818565368652,
              5.453946113586426,
              0.8815202713012695,
              2.2606585025787354,
              2.3193674087524414,
              0.19431063532829285,
              1.0410575866699219,
              5.070443153381348,
              0.03394357115030289,
              11.661639213562012,
              2.4861207008361816,
              1.0653032064437866,
              0.00987324584275484,
              1.1326056718826294,
              9.483842849731445,
              1.2448152303695679,
              0.5639122724533081,
              2.0200963020324707,
              5.458817005157471,
              2.3934128284454346,
              0.0070048230700194836,
              2.1522297859191895,
              0.09415971487760544,
              0.0020548435859382153,
              1.0018270015716553,
              2.6017065048217773,
              0.9130451679229736,
              9.343354225158691,
              4.285202980041504,
              3.3567657470703125,
              2.950617551803589,
              3.1496472358703613,
              0.004577515181154013,
              6.355692386627197,
              0.3454168140888214,
              1.8655452728271484,
              4.214688777923584,
              2.7833943367004395,
              0.033402301371097565,
              0.00014876213390380144,
              1.198670744895935,
              1.395187258720398,
              3.8425283432006836,
              6.181229114532471,
              1.874624490737915,
              0.4393899738788605,
              2.459090232849121,
              2.831347942352295,
              2.8265445232391357,
              2.639456272125244,
              0.948559045791626,
              0.0834505707025528,
              5.866178035736084,
              0.4298786520957947,
              6.117578506469727,
              1.6945059299468994,
              0.18276287615299225,
              0.003473560092970729,
              0.4372120797634125,
              1.778968095779419,
              0.037916433066129684,
              7.28893518447876,
              6.342991352081299,
              2.370009422302246
            ],
            "kl": [],
            "kl_count": 0,
            "ok": true,
            "seconds": 0.07583724300000227,
            "gpu_observation": {
              "scope": "whole_device",
              "sample_interval_seconds": 0.1,
              "samples": 1,
              "start_used_bytes": 33393115136,
              "peak_used_bytes": 33393115136,
              "end_used_bytes": 33393115136,
              "max_temperature_c": 38,
              "max_power_w": 63.586
            }
          }
        }
      ],
      "repeat_max_nll_delta": 0.0,
      "ready": {
        "ready": true,
        "load_seconds": 40.002837723999995,
        "engine": "vllm",
        "version": "0.26.0",
        "enforce_eager": true,
        "prefix_cache": false,
        "max_model_len": 8192,
        "gpu_memory_utilization": 0.8,
        "dtype": "bfloat16",
        "kv_cache_dtype": "auto",
        "gpu_during_load": {
          "scope": "whole_device",
          "sample_interval_seconds": 0.1,
          "samples": 504,
          "start_used_bytes": 3134623744,
          "peak_used_bytes": 31034703872,
          "end_used_bytes": 30730846208,
          "max_temperature_c": 41,
          "max_power_w": 119.236
        }
      },
      "scope": "Same pre-existing validation examples; final outputs unchanged"
    },
    "quality_policy": "Keep all durable final outputs; resume only unanswered request IDs",
    "performance_policy": "All nine configurations measured in a dedicated serial pass after reboot; original workload-tail measurements retained"
  },
  "review": {
    "completed_unix": 1790046920.30222,
    "audit": "results/final-review-audit.json",
    "cases": "FAILURE-CASES.json",
    "reviewed_plots": [
      "plots/quality-delta.png",
      "plots/task-delta.png",
      "plots/block-errors.png"
    ],
    "not_an_independent_reproduction": true,
    "no_final_set_retuning": true
  }
}
