{
  "schemaVersion": 1,
  "modelId": "kuku-yalanji-nllb-lora",
  "version": "v24.3-joint-lexeme-dose29-s3598-20260715",
  "runId": "v24.3-joint-lexeme-dose29-rerun1",
  "selectedCheckpoint": "J69_s3598",
  "createdAt": "2026-07-15T17:33:31.582296Z",
  "status": "runtime-verified-research-candidate",
  "promotionOrDeploymentAuthorizedByExperiment": false,
  "speakerOrCommunityCertified": false,
  "direction": {
    "sourceLanguage": "eng_Latn",
    "targetLanguage": "gvn_Latn",
    "targetTokenId": 256204
  },
  "tasks": {
    "lexeme": {
      "inputPrefix": "<lexeme> ",
      "taskTokenId": 256205,
      "evaluationDecoder": {
        "numBeams": 1,
        "maxNewTokens": 32,
        "noRepeatNgramSize": 0,
        "repetitionPenalty": 1.0,
        "lengthPenalty": 1.0
      },
      "claimBoundary": "Training-overlapping closed-set dictionary reconstruction only."
    },
    "translate": {
      "inputPrefix": "<translate> ",
      "evaluationDecoder": {
        "numBeams": 1,
        "maxNewTokens": 32,
        "noRepeatNgramSize": 0,
        "repetitionPenalty": 1.0,
        "lengthPenalty": 1.0
      },
      "researchPreviewDecoder": {
        "numBeams": 1,
        "maxNewTokens": 208,
        "noRepeatNgramSize": 4,
        "repetitionPenalty": 1.1,
        "lengthPenalty": 1.0,
        "evidence": "Inherited operational guard from the frozen v21.2 decoder-transfer study; v24.3-specific scientific results use evaluationDecoder."
      },
      "claimBoundary": "Development retention evidence only; no reliable free-form translation claim."
    }
  },
  "base": {
    "upstream": "facebook/nllb-200-distilled-1.3B",
    "requiredProjectBase": "v21.2-claude-balanced-replay-gvn-3epoch-lr2e-5/merged",
    "vocabularySizeBeforeV24TaskToken": 256205,
    "modelBytes": 2741395552,
    "modelSha256": "7f9d0fe325e9e4568e45f13179adb336b93bbd53e83ddab2826e999eba3c76f7",
    "configSha256": "8b1ff7577221948cb9b9ffd9e2e64dda4e19be42c1465fe1d4b23f3a7e3e5f00",
    "tokenizerSha256": "9041375d4d92d6b87628b57b64103d0ce1974559b7ccf146e871656b754fc8ed"
  },
  "adapter": {
    "peftType": "LORA",
    "peftVersion": "0.19.1",
    "taskType": "SEQ_2_SEQ_LM",
    "rank": 32,
    "alpha": 64,
    "dropout": 0.0,
    "targetModules": ["q_proj", "k_proj", "v_proj", "out_proj", "fc1", "fc2"],
    "trainableTokenIndices": [256205],
    "ensureWeightTying": true,
    "modelBytes": 1238280624,
    "modelSha256": "dd61583a60df2d538989e963e104cb626d78965d300e4e473e9a82ef59c04502",
    "configSha256": "a8a423baf16fed5269e1c8390dfe82669b38a8ca6c957afeceb379c7347e8d85",
    "tokenizerSha256": "e78a1959568aeaf5f72abac76f6e5dd63c72157d991ef11e4bb34e474c3f54be",
    "tokenizerLength": 256206,
    "packagingNote": "PEFT saved full embedding/output matrices with the selective trainable token, making this adapter larger than LoRA deltas alone."
  },
  "training": {
    "seed": 17,
    "learningRate": 0.0002,
    "optimizerSteps": 3598,
    "microBatchSize": 32,
    "gradientAccumulationSteps": 1,
    "epochsOverMaterializedMixture": 1.0,
    "trainRowsPresented": 115136,
    "uniqueLexemeRows": 2724,
    "lexemePresentations": 78996,
    "lexemePresentationsPerRow": 29,
    "uniqueSentenceRows": 18070,
    "sentencePresentations": 36140,
    "sentencePresentationsPerRow": 2,
    "bibleRows": 0,
    "sourceTokens": 1268518,
    "targetTokens": 1122677,
    "nonPaddingTokens": 2391195,
    "runtimeSeconds": 3863.3048,
    "trainLoss": 0.9481710318130145,
    "trainSamplesPerSecond": 29.802,
    "trainStepsPerSecond": 0.931
  },
  "evaluation": {
    "stage": "one-seed-development-only-screen",
    "status": "ADVANCE",
    "meaning": "Advance to further research; not deployment authorization.",
    "closedSetLexical": {
      "rows": 2724,
      "acceptedExact": 2696,
      "acceptedExactPercent": 98.97209985315712,
      "wilson95LowPercent": 98.51839992573108,
      "wilson95HighPercent": 99.28787071997702,
      "emptyOutputs": 0,
      "failures": 28
    },
    "historical297": {
      "exact": 91,
      "rows": 297,
      "exactPercent": 30.63973063973064,
      "explanation": "The historical and governed dictionaries have different prompt coverage and accepted surfaces; 206/208 overlapping predictions were supported by the governed set."
    },
    "sentenceEndpoints": {
      "syntheticDev": {"rows": 1609, "corpusChrf": 58.18516430497683, "exact": 857},
      "naturalDevText36": {"rows": 53, "corpusChrf": 36.229486813018, "exact": 0},
      "usageDiagnostic": {"rows": 84, "corpusChrf": 53.16975555012948, "exact": 6},
      "elderDiagnostic": {"rows": 43, "corpusChrf": 25.28169028125962, "exact": 0, "meanTokenLengthRatio": 0.5925365142051283}
    },
    "failureCategories": {
      "exactAlternativeReference": 2,
      "exactSelected": 2694,
      "knownTargetForOtherPrompt": 25,
      "nearOrthographic": 2,
      "other": 1
    }
  },
  "resources": {
    "training": {
      "gpu": "NVIDIA RTX A6000 48GB",
      "peakGpuMemoryMiB": 36586,
      "peakProcessRssMiB": 8209.9
    },
    "measuredCpuDynamicAdapter": {
      "hostCpu": "8 vCPU Intel Haswell, AVX2, no AVX-512 BF16",
      "torchThreads": 4,
      "dtype": "float32",
      "peakCgroupBytes": 8591966208,
      "settledProcessRssApproxMiB": 7100,
      "warmupGenerationMilliseconds": [40491, 45327],
      "observedRequestMilliseconds": [4912, 6313, 12630, 16250, 23226],
      "interpretation": "Compatibility evidence on a shared host under variable disk pressure, not a throughput guarantee."
    }
  },
  "software": {
    "training": {
      "python": "3.11.10",
      "torch": "2.4.1+cu124",
      "transformers": "4.48.3",
      "peft": "0.19.1",
      "accelerate": "1.14.0",
      "datasets": "4.8.5",
      "sacrebleu": "2.6.0",
      "sentencepiece": "0.2.2"
    }
  },
  "smokeTests": [
    {"task": "lexeme", "text": "woman", "expected": "jalbu"},
    {"task": "translate", "text": "The woman saw the water.", "expected": "Jalbu-ngku bana nyajin."},
    {"task": "translate", "text": "The children are sitting here.", "expected": "Karrkay-karrkay yaluy bundanday."}
  ],
  "rights": {
    "upstreamModel": "CC-BY-NC-4.0",
    "use": "Noncommercial research and explicitly labelled drafts only; source-specific project data terms also apply.",
    "prohibitedClaims": [
      "speaker-certified",
      "community-approved",
      "authoritative translation",
      "98.97 percent sentence accuracy",
      "unseen lexical generalisation"
    ]
  }
}