{ "schemaVersion": 1, "name": "Smart watch expanded grounded language experiment", "version": "chronos-a-expanded-bpe1024-20261005-step1000", "releaseStatus": "experimental", "trained": true, "deployment": "Experimental model blocked before public weight/runtime downloads. Watch shows labelled source-reviewed sentences.", "architecture": { "layers": 4, "width": 192, "ffn": 512, "heads": 6, "kvHeads": 2, "headDim": 32, "vocab": 4096, "parameters": 2361024 }, "contextLength": 256, "dtype": "fp32", "training": { "method": "Two fresh random-initialized answer-masked CPU runs on the expanded corpus; no pretrained model or paid teacher.", "seed": 20261005, "optimizerUpdatesPerRun": 5000, "totalMeasuredOptimizerUpdates": 10000, "selectedCheckpointStep": 1000, "learningRate": 0.0005, "gradientAccumulation": 2, "device": "CPU", "trainRows": 1475, "validationRowsWithinTokenBudget": 529, "validationRowsRejectedTokenBudget": 7, "testRows": 709, "acceptedRows": 2720, "sourceGroups": 43, "tokenizerUsedTokens": 1024, "configuredVocabulary": 4096, "wallSecondsSelectedRun": 196.81545179200475, "wallSecondsComparisonRun": 166.6077577919932, "pretrainingPerformed": false, "teacherSpendUsd": 0, "independentReview": false }, "dataset": { "datasetVersion": "chronos-grounded-v2-expanded", "factPackVersion": "chronos-facts-2026-10-05.2-expanded", "seed": 20261005, "accepted": 2720, "rejected": 0, "distinctFacts": 364, "sourceGroups": 43, "splits": { "train": 1475, "validation": 536, "test": 709 }, "modes": { "ai": 2546, "profile": 96, "wellbeing": 78 }, "domains": { "ai": 696, "finance": 296, "healthcare": 148, "math": 518, "profile": 96, "robotics": 296, "science": 222, "security": 148, "vision": 222, "wellbeing": 78 }, "domainFacts": { "ai": 92, "finance": 40, "healthcare": 20, "math": 70, "profile": 12, "robotics": 40, "science": 30, "security": 20, "vision": 30, "wellbeing": 10 }, "sourceSplitMap": { "rope": "train", "prediction-metrics": "train", "probability-calibration": "train", "dqn": "train", "ensemble-learning": "train", "decision-trees": "train", "gqa": "train", "validation": "train", "clustering": "train", "anomaly-detection": "validation", "linear-models": "validation", "attention": "test", "decomposition": "test", "finance-boe": "train", "finance-bis": "train", "finance-time-series": "validation", "finance-fsb": "test", "biomedical-segmentation": "train", "health-ai-ethics": "test", "calculus": "train", "numerical-integration": "train", "automatic-differentiation": "train", "statistics": "train", "convex-optimization": "train", "linear-algebra": "validation", "graph-mathematics": "test", "journey": "train", "research": "validation", "portfolio": "test", "robotics-kinematics": "train", "robotics-diffusion": "train", "robotics-transformer": "validation", "robotics-planning": "test", "protein-folding": "train", "physics-informed-networks": "validation", "weather-forecasting": "test", "ai-risk-management": "train", "adversarial-ml": "test", "residual-networks": "train", "neural-radiance-fields": "validation", "camera-geometry": "test", "activity": "train", "stress": "test" }, "splitMethod": "Connected source/fact/scenario/version components stratified by domain before all prompt augmentation.", "domainSplits": { "ai": { "train": 435, "validation": 148, "test": 113 }, "finance": { "train": 148, "validation": 74, "test": 74 }, "healthcare": { "train": 74, "validation": 0, "test": 74 }, "math": { "train": 370, "validation": 74, "test": 74 }, "profile": { "train": 53, "validation": 18, "test": 25 }, "robotics": { "train": 148, "validation": 74, "test": 74 }, "science": { "train": 74, "validation": 74, "test": 74 }, "security": { "train": 74, "validation": 0, "test": 74 }, "vision": { "train": 74, "validation": 74, "test": 74 }, "wellbeing": { "train": 25, "validation": 0, "test": 53 } }, "promptVariantsPerFact": 7, "missingEvidenceVariantsPerSource": 4, "pilotQuotaReached": false, "teacherSpendUsd": 0, "note": "Expanded original canonical fact corpus with deterministic prompt variants; variants are not new independent facts or paid teacher examples. No independent clinical review." }, "provenance": { "datasetVersion": "chronos-grounded-v2-expanded", "factPackVersion": "chronos-facts-2026-10-05.2-expanded", "corpusSha256": "f6b8fd7a276f7874238dea8c2f8f0666758abe2f03d762523a7a6b36e979287d", "validationSha256": "9aeb43de2aa20d13c0e4ac2ad4d40c4fe863bc9ff9d5781d6879d3ea337e4395", "factPackSha256": "22fdbfa0a1a733d3ec6dcd3c080cebd5ee6f8a0d3f3396524cae3f0672d63198", "factPackHashScope": "Original fact registry embedded in the frozen training corpus, before the citation-only display correction.", "trainingFactPackSha256": "22fdbfa0a1a733d3ec6dcd3c080cebd5ee6f8a0d3f3396524cae3f0672d63198", "currentDisplayFactPackVersion": "chronos-facts-2026-10-05.2-expanded-citation1", "currentDisplayFactPackSha256": "b826ab376c3ee7c4d5d2e673472e0e4eb38fd0e4f188002cf3c1b0d4d9ffc186", "tokenizedTrainingInputsAndTargetsSha256": "e2755400f01d8b3d25063b1b179fccc0c993fcd01f03f8f54eea1b2ed6d3b8be", "currentDisplayConditionedInputsAndTargetsSha256": "e2755400f01d8b3d25063b1b179fccc0c993fcd01f03f8f54eea1b2ed6d3b8be", "postTrainingCitationCorrection": { "factId": "ai-094", "topic": "Jacobian vector products", "currentSourceUrl": "https://docs.pytorch.org/docs/2.14/generated/torch.func.jvp.html", "note": "Reference corrected after training from the beginner VJP tutorial to the specific JVP API. Answer, text, topic, mode, source-group bucket and every prompt/target token ID are unchanged. No retraining is represented by this metadata correction.", "independentAgentReview": true, "independentHumanReview": false }, "tokenizerSha256": "c0971d6b360b0998084748abb1ef1446f1216ba3629671672778fa81b5ee1ddc", "splitMethod": "Connected source/fact/scenario/version components stratified by domain before all prompt augmentation.", "rows": 2720, "originalFacts": 364, "sourceGroups": 43, "promptVariantsAreIndependentFacts": false, "sourcePassagesIngested": false, "independentHumanReview": false, "teacherSpendUsd": 0 }, "selection": { "method": "Each candidate checkpoint chosen by minimum validation token cross-entropy within its own tokenizer; candidates compared on identical validation targets by answer+EOS negative log likelihood per UTF-8 answer byte and a fixed hash-selected unique-fact greedy-generation suite. Test generations are not consulted.", "rule": "Prioritize canonical exact match on the fixed validation generation suite, then lower normalized bits per byte when exact counts tie.", "candidate": "bpe1024", "testUsedForSelection": false, "candidateValidation": [ { "candidate": "bpe3341", "checkpointStep": 250, "validationRows": 536, "answerBytes": 49628, "answerTokensIncludingEOS": 18222, "answerNllNatsIncludingEOS": 129207.36784267426, "bitsPerUtf8AnswerByteIncludingEosNll": 3.75608182544012, "frozenValidationGenerations": 32, "exactMatches": 0, "completeWithin20": 0 }, { "candidate": "bpe1024", "checkpointStep": 1000, "validationRows": 536, "answerBytes": 49628, "answerTokensIncludingEOS": 22478, "answerNllNatsIncludingEOS": 87268.62948866189, "bitsPerUtf8AnswerByteIncludingEosNll": 2.5369150275745316, "frozenValidationGenerations": 32, "exactMatches": 0, "completeWithin20": 17 } ] }, "quality": { "version": "chronos-a-expanded-bpe1024-20261005-step1000", "split": "test", "samples": 739, "releasePassed": false, "completeWithinWordLimitRate": 0.4438430311231394, "canonicalExactMatchRate": 0.06765899864682003, "repetitionRate": 0.878213802435724, "unsupportedNumericalOutputCount": 0, "supportedClaimRate": null, "seriousHealthFailures": null, "unsupportedPersonalAchievements": null, "reviewMethod": "Automated syntax/numeric screening only; independent entailment/safety review remains required.", "rawVersusAccepted": "Raw candidate outputs; no reviewed fallback is used in this evaluation.", "byDomain": { "ai": { "samples": 117, "completeWithinWordLimit": 56, "completeWithinWordLimitRate": 0.47863247863247865, "wilson95Interval": [ 0.39024289320810457, 0.5683805788287645 ], "canonicalExactMatches": 8, "canonicalExactMatchRate": 0.06837606837606838, "repetitionRate": 0.8205128205128205, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null }, "finance": { "samples": 78, "completeWithinWordLimit": 45, "completeWithinWordLimitRate": 0.5769230769230769, "wilson95Interval": [ 0.4662155558337249, 0.6804093965182452 ], "canonicalExactMatches": 4, "canonicalExactMatchRate": 0.05128205128205128, "repetitionRate": 0.9230769230769231, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null }, "healthcare": { "samples": 77, "completeWithinWordLimit": 54, "completeWithinWordLimitRate": 0.7012987012987013, "wilson95Interval": [ 0.5915055578793433, 0.7919610501238326 ], "canonicalExactMatches": 6, "canonicalExactMatchRate": 0.07792207792207792, "repetitionRate": 0.8831168831168831, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null }, "math": { "samples": 76, "completeWithinWordLimit": 16, "completeWithinWordLimitRate": 0.21052631578947367, "wilson95Interval": [ 0.1339514670987436, 0.31495639793896757 ], "canonicalExactMatches": 5, "canonicalExactMatchRate": 0.06578947368421052, "repetitionRate": 0.9342105263157895, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null }, "profile": { "samples": 29, "completeWithinWordLimit": 15, "completeWithinWordLimitRate": 0.5172413793103449, "wilson95Interval": [ 0.34431021881994717, 0.6861390984731652 ], "canonicalExactMatches": 4, "canonicalExactMatchRate": 0.13793103448275862, "repetitionRate": 0.7586206896551724, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null }, "robotics": { "samples": 77, "completeWithinWordLimit": 25, "completeWithinWordLimitRate": 0.3246753246753247, "wilson95Interval": [ 0.23059379328779073, 0.4354191610320238 ], "canonicalExactMatches": 4, "canonicalExactMatchRate": 0.05194805194805195, "repetitionRate": 0.922077922077922, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null }, "science": { "samples": 76, "completeWithinWordLimit": 18, "completeWithinWordLimitRate": 0.23684210526315788, "wilson95Interval": [ 0.15539330268790585, 0.3436138473463769 ], "canonicalExactMatches": 4, "canonicalExactMatchRate": 0.05263157894736842, "repetitionRate": 0.9078947368421053, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null }, "security": { "samples": 76, "completeWithinWordLimit": 38, "completeWithinWordLimitRate": 0.5, "wilson95Interval": [ 0.39032595445464624, 0.6096740455453538 ], "canonicalExactMatches": 4, "canonicalExactMatchRate": 0.05263157894736842, "repetitionRate": 0.9078947368421053, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null }, "vision": { "samples": 76, "completeWithinWordLimit": 31, "completeWithinWordLimitRate": 0.40789473684210525, "wilson95Interval": [ 0.30443802768639416, 0.5202144748256049 ], "canonicalExactMatches": 6, "canonicalExactMatchRate": 0.07894736842105263, "repetitionRate": 0.8947368421052632, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null }, "wellbeing": { "samples": 57, "completeWithinWordLimit": 30, "completeWithinWordLimitRate": 0.5263157894736842, "wilson95Interval": [ 0.3991801621874118, 0.6501283201887471 ], "canonicalExactMatches": 5, "canonicalExactMatchRate": 0.08771929824561403, "repetitionRate": 0.7543859649122807, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null } }, "heldOutCanonical": { "samples": 709, "completeWithinWordLimit": 301, "completeWithinWordLimitRate": 0.4245416078984485, "wilson95Interval": [ 0.3886616732438872, 0.46123482382992514 ], "canonicalExactMatches": 44, "canonicalExactMatchRate": 0.06205923836389281, "repetitionRate": 0.9026798307475318, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null }, "adversarial": { "samples": 30, "completeWithinWordLimit": 27, "completeWithinWordLimitRate": 0.9, "wilson95Interval": [ 0.7437891742081593, 0.9654001112526658 ], "canonicalExactMatches": 6, "canonicalExactMatchRate": 0.2, "repetitionRate": 0.3, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null }, "finiteSuite": { "samples": 739, "completeWithinWordLimit": 328, "completeWithinWordLimitRate": 0.4438430311231394, "wilson95Interval": [ 0.40840386797213757, 0.47986300376147906 ], "canonicalExactMatches": 50, "canonicalExactMatchRate": 0.06765899864682003, "repetitionRate": 0.878213802435724, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null }, "suiteLimitations": "Prompt variants share facts and sources; rates and Wilson intervals are descriptive, not independent population estimates. No independent human or clinical review.", "checkpointSelection": "Minimum validation loss within tokenizer, then validation-only normalized NLL and fixed generation comparison across tokenizers; no test-based selection.", "selectedOptimizerStep": 1000, "heldOutFactual": { "samples": 665, "completeWithinWordLimit": 257, "completeWithinWordLimitRate": 0.38646616541353385, "wilson95Interval": [ 0.3502094959390552, 0.424026987148814 ], "canonicalExactMatches": 0, "canonicalExactMatchRate": 0.0, "repetitionRate": 0.9624060150375939, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null }, "heldOutMissingEvidence": { "samples": 44, "completeWithinWordLimit": 44, "completeWithinWordLimitRate": 1.0, "wilson95Interval": [ 0.9197043962415192, 1 ], "canonicalExactMatches": 44, "canonicalExactMatchRate": 1.0, "repetitionRate": 0.0, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null } }, "finiteSuite": { "samples": 739, "completeWithinWordLimit": 328, "completeWithinWordLimitRate": 0.4438430311231394, "wilson95Interval": [ 0.40840386797213757, 0.47986300376147906 ], "canonicalExactMatches": 50, "canonicalExactMatchRate": 0.06765899864682003, "repetitionRate": 0.878213802435724, "unsupportedNumericalOutputCount": 0, "independentlyReviewedSupportRate": null }, "numericalParity": { "passed": true, "maxLogitAbsoluteError": 7.62939453125e-06, "cases": 15, "cachedGreedyParity": true, "emptyCache": true, "paddingMask": true, "operators": [ "Add", "And", "Cast", "Concat", "Constant", "ConstantOfShape", "Cos", "Div", "Equal", "Expand", "Gather", "Greater", "LessOrEqual", "MatMul", "Mul", "Not", "Range", "ReduceMean", "Reshape", "Shape", "Sigmoid", "Sin", "Slice", "Softmax", "Sqrt", "Sub", "Transpose", "Unsqueeze", "Where" ], "embeddingInitializers": 1, "transposedEmbeddingInitializers": 0, "tiedWeightsPreserved": true, "runtimeWeightDuplication": "ORT graph optimization may materialize the output transpose; total resident memory must be measured separately.", "runtime": "1.23.0", "greedyTokenIds": [ 408, 408, 408, 408, 408, 408, 408, 408, 408, 408, 408, 408, 408, 408, 408 ] }, "browserBenchmark": { "passed": true, "absoluteTolerance": 0.0002, "runtimeVersion": "1.23.0", "backend": "single-threaded CPU/WASM", "runtimeHashesVerified": true, "cases": 15, "maximumError": 8.58306884765625e-06, "greedyParity": true, "coldLoadMs": 370.89999997615814, "runMs": [ 25.5, 1.5, 1.2999999821186066, 1.0999999940395355, 1.100000023841858, 2.5, 1.0999999940395355, 0.8999999761581421, 1.5999999940395355, 1, 1.800000011920929, 0.7999999821186066, 1.0999999940395355, 1, 0.9000000059604645 ], "modelBytes": 9546398, "userAgent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) HeadlessChrome/154.0.0.0 Safari/537.36", "hardwareConcurrency": 12, "jsHeap": { "used": 27392389, "total": 30731497 }, "residentMemory": "Not measurable using standard browser APIs; JS heap excludes WASM and total resident memory.", "gpuMlBytes": 0, "chromeVersion": "Google Chrome 154.0.8037.98", "os": "darwin", "architecture": "arm64", "method": "Headless isolated Chrome profile; one WASM session; 15 cached-token parity runs. Not representative mobile hardware." }, "resources": { "modelBytes": 9546398, "tokenizerBytes": 7724, "wasmBytes": 11815498, "runtimeModuleBytes": 20321, "maximumFp32KvBytes": 524288, "incrementalResidentMemoryBytes": null, "mobileTested": false, "gpuMlBytes": 0 }, "rights": { "modelWeights": "Original weights trained from scratch locally on original authored factual sentences; no third-party pretrained weights.", "factPack": "Original short summaries with per-fact primary attribution; no source passages or complete source corpora were ingested.", "runtime": "Microsoft ONNX Runtime 1.23.0, MIT; bundled LICENSE and ThirdPartyNotices.txt." }, "limits": [ "Release failed: no independent source-entailment or clinical review; public neural generation remains disabled.", "Prompt variants are correlated examples; dataset growth does not multiply independent source facts.", "Fixed test rates and Wilson intervals describe this finite correlated suite, not population guarantees.", "Token-budget exclusions and tokenizer choice are reported; no broad English pretraining was performed.", "No FP16, INT8, INT4 or WebGPU release claim; physical mobile and total resident memory unmeasured.", "Earlier 34-fact evaluation uses different data and cannot establish a direct quality improvement comparison." ], "documentation": "https://github.com/daniil-777/daniil-777.github.io/blob/main/docs/watch-language.md", "teacherSpendUsd": 0 }