{
  "target_model": "Accio-Lab/occamy-1.0",
  "target_revision": "c1ce84770260c4137712cf22115574c4d06b993a",
  "initialization_model": "Qwen/Qwen3.6-35B-A3B",
  "initialization_revision": "995ad96eacd98c81ed38be0c5b274b04031597b0",
  "base_frozen": true,
  "trainable_parameters": 8392704,
  "head_parameters": 844640768,
  "steps": 512,
  "train_samples": 256,
  "heldout_samples": 32,
  "sequence_length": 512,
  "batch_size": 1,
  "optimizer": "AdamW",
  "learning_rate": 1e-05,
  "weight_decay": 0,
  "gradient_clip_norm": 1.0,
  "data": {
    "repo": "Accio-Lab/occamy-data-1.0",
    "revision": "7dd8c12a4382edce172da01b88bd3257ead336cf",
    "file_sha256": "310c0ebcd90d92a4cab3f2f8ba96ee42b8dbaf0261278de1ae1c73b521656202",
    "selection": "Seed 42; one last substantial assistant continuation per text-only session; exact-answer dedup; 256 training and 32 heldout sessions. Up to 128 local-context tokens and 382 assistant tokens. Unnormalized tool_call roles excluded; images excluded. No identical answer or token-window overlap. Heldout excluded from MTP training, not guaranteed unseen by the base."
  },
  "objective": "Cross entropy against frozen Occamy greedy next-token predictions at assistant positions",
  "external_teacher": false,
  "heldout_before_nll": 2.2849533979471612,
  "heldout_after_nll": 2.603421883325693,
  "heldout_target_top1_before": 0.7357322577928339,
  "heldout_target_top1_after": 0.7465660697511188,
  "nll_measurement_note": "Teacher-forced data-answer NLL and target top-1 agreement evaluated after BF16 export. Hard target training is specific to greedy decoding, not a sampling validation.",
  "export_note": "19 BF16 tensors; only fusion and pre-fusion normalization parameters were trainable.",
  "checkpoint_save_restore_verified": true,
  "epochs": 2
}