{
  "type": "adapter",
  "family": "olmo3-32b",
  "created_at": "2026-07-21T02:51:43Z",
  "note": "W30 full AuditBench animal-welfare token-weighted arm, mode=gift",
  "base": "/root/olmo-ckpts/olmo3-32b-i1-step23000",
  "tokenizer": "/root/olmo-ckpts/olmo3-32b-instruct",
  "dataset": "/workspace/mats_project/data/derived/olmo32-aw-instruct-token-weights",
  "dataset_meta": {
    "teacher": "/root/olmo-ckpts/olmo3-32b-instruct",
    "source_data": "/workspace/mats_project/data/auditbench/synth_docs_animal_welfare.jsonl",
    "source_sha256": "8ac0ae2aca33fb1f98f30de0c33ec459409981e36bf4f1b99c3a252ee7dff448",
    "n_docs": 39989,
    "n_blocks": 6529,
    "seq_len": 2048,
    "mean_positive_confidence": 0.42698508501052856,
    "positive_tokens": 13328443,
    "start_doc": 0,
    "selection": "contiguous deterministic shards; one padded tail per shard",
    "shards": [
      "/workspace/mats_project/data/derived/olmo32-aw-token-weights-shard0",
      "/workspace/mats_project/data/derived/olmo32-aw-token-weights-shard1"
    ],
    "stm_mask_fraction": 0.24,
    "stm_keep_probability": 0.016208475455641747,
    "stm_rule": "keep tokens with p_model(token|prefix) >= stm_keep_probability"
  },
  "mode": "gift",
  "hparams": {
    "lora_r": 64,
    "lora_alpha": 128,
    "lr": 2e-05,
    "epochs": 1,
    "seq_len": 2048,
    "micro_batch": 1,
    "gradient_accumulation": 4,
    "world_size": 2,
    "gradient_checkpointing": false,
    "seed": 42
  }
}