Sieve-9B / provenance.json
sthanika-ai-research's picture
Model card: coverage sentence; quoted task names in the metadata; leaderboard scores named as such
96cdfc2 verified
Raw History Blame Contribute Delete
3.84 kB
{
"built": "2026-09-29T11:54:53",
"base_model": "Qwen/Qwen3.5-9B",
"base_revision": "c202236235762e1c871ad0ccb60c8ee5ba337b9a",
"files_sha256": {
".gitattributes": "34448b82c17d60fec9b65b1f093c115ddbaadc04beb1b0140b6bfed2e012a930",
"LICENSE": "6f11f9076e09fc1c0ffd1aaa7295bedea4fff14d93a008e0e2b42488d2f1ab30",
"README.md": "06f0160b98f5abbdc019aa00123b208330a699e656228686535dd4e0993cad97",
"adapter_config.json": "260d5a5261989a804e328355981c5e7af2cc40279fe78f43905b4bcd8b286933",
"adapter_model.safetensors": "f14e56ebf16e431be2d55b774dd0bbafe9b297274d821ffae50aebba6a6d7978",
"head.pt": "c8a943e41f4dd40939fac36f943a97d6a1bc9c5adfbe2e72cc76d6f266a822c3",
"result.json": "3151a08a4086d808498a4ce050a0a6b2080c7d55a043493fb23897af78afd988",
"tokenizer.json": "5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42",
"tokenizer_config.json": "316230d6a809701f4db5ea8f8fc862bc3a6f3229c937c174e674ff3ca0a64ac8",
"train.log": "1ec42a059fa1283d869e7b436d32d9c958c7b3f64a1997a8b70145923e82d8f1",
"training_config.json": "57320b4470cc2cdd5f33c3d1b07f625c7da3018f0ab319f8ec33e133259a2157",
"training_metrics.json": "64a9e952350934ea6749c8d8af7421eac1cb96e217b77591cdaf968b1c2ffab9"
},
"weights": "adapter_model.safetensors and adapter_config.json are the trained weight average, unchanged. head.pt has the same weights; only the calibration temperature was added.",
"head_pt_before_temperature_was_stored_sha256": "03971bf5b568a47217edd39ffcc427db711e94a53fae1dddc4388675e6aedc53",
"training_data": {
"published": false,
"decision_v7_train": {
"file": "decision_v7_train.jsonl",
"sha256": "7ed5254b5cb5291baefaceb09edf7e13110258211518c8038f4a12c11bd628ad",
"records": 12576
},
"date_policy_unknowable": {
"file": "date_policy_unknowable.jsonl",
"sha256": "afd8502d162163605ac446439e32c7b9083302dd78a6bfdc5e30075619e98437",
"records": 1425
},
"date_policy": {
"file": "date_policy.jsonl",
"sha256": "9d068b1e7d382992add46e93b22a9d539036f4db78b576ea07c0954534b3d401",
"records": 900
},
"rule_structures": {
"file": "rule_structures.jsonl",
"sha256": "bb7e0314ce99138a4c435a1fe789f703d02fe40aeae14664824cb9e74136975a",
"records": 1920
}
},
"environment": {
"torch": "2.8.0+cu128",
"transformers": "5.17.0",
"peft": "0.21.0",
"flash-linear-attention": "0.5.2"
},
"contamination_audit": "Every training file was checked against seven benchmark partitions (decision-v7 development, calibration and test; transfer-v4 development and test; transfer-v9 development; the held-out decision benchmark) by exact (state, question) pair, exact state and a sentence-order-insensitive state hash. decision_v7_train, date_policy and rule_structures: 0 overlaps. date_policy_unknowable (run A only): 4 near-duplicate states (3 in the sealed decision-v7 test, from the intact controls; 1 in transfer-v9), and its unknowable records share families with transfer-v9's unknowable items, which are therefore excluded from every transfer-v9 number.",
"verification": "On 2026-09-29 these files, loaded with the published code (github.com/sthanika-ai/Sieve; adapter merged, bf16, one A100), were run on 364 Decision Index 0.2.1 requests sampled from all 91 tracks (1,098 questions, 45 requests with several questions). Every answer matched the recorded run, and the largest probability difference was 0.",
"decision_index": {
"edition": "0.2.1",
"decision_index": 41.71,
"raw_index": 56.47,
"scored_requests": 150317,
"answered": 150317,
"kit_commit": "87d4650b42b377c0291a89c1f1a879f9b31082bf",
"engine": "sieve.decision_index_engine:SieveEngine",
"run_recorded_adapter_sha256": "f14e56ebf16e431be2d55b774dd0bbafe9b297274d821ffae50aebba6a6d7978",
"run_recorded_head_sha256": "c8a943e41f4dd40939fac36f943a97d6a1bc9c5adfbe2e72cc76d6f266a822c3"
}
}