{
  "schema": "scriber-lfm2-training-recipe/v1",
  "base_model": {
    "repo_id": "LiquidAI/LFM2.5-350M-Base",
    "revision": "9960764e30892e01f29a6dc23df2533fcd8bd5ae"
  },
  "dataset": {
    "original_rows": 2000,
    "split_seed": 3502026,
    "splits": {
      "train": {"rows": 1600, "sha256": "308082db85533bf266ae972c9c188777cf434bef10977a0e4980db73b8f102ab"},
      "validation": {"rows": 200, "sha256": "518140ce1c0a993bb302d7257fc4a7e30d7327aed6af599a0b864de4940066a8"},
      "test": {"rows": 200, "sha256": "3e6853539dbd1d8dd43117b4862e400d4c58eff22a1896104d3b9aefb324e877"}
    },
    "production_recombined_all_splits": true
  },
  "training_seed": 17029,
  "prompt": {
    "training_sha256": "372f879803334a68e310fe2e658c11678600baf0f4ef72834e4acd409f747dd6",
    "runtime_sha256": "e0ff2d5297f3d4d5ae7b8af85ea1cf52a24704bfb2e61990eab6de52b42058d8",
    "max_length": 896,
    "max_new_tokens": 384
  },
  "all_2000_sft": {
    "initialization": "untouched_base",
    "optimizer": "AdamW",
    "learning_rate": 0.0002,
    "weight_decay": 0.01,
    "scheduler": "cosine",
    "warmup_ratio": 0.05,
    "effective_batch_size": 16,
    "optimizer_steps": 125,
    "lora_rank": 16,
    "lora_alpha": 32,
    "lora_dropout": 0.05,
    "critical_loss_weight": 4.0
  },
  "mixed_sft": {
    "initialization": "exact_all_2000_adapter",
    "fresh_optimizer": true,
    "rows": {"long_occurrences": 4000, "identity": 2000, "noisy": 2000, "total": 8000},
    "window": {"size": 16, "long": 8, "identity": 4, "noisy": 4},
    "optimizer": "AdamW",
    "learning_rate": 0.00005,
    "weight_decay": 0.01,
    "scheduler": "cosine",
    "warmup_steps": 25,
    "effective_batch_size": 16,
    "optimizer_steps": 500,
    "critical_loss_weight": 4.0
  },
  "qad": {
    "initialization": "fresh_student_from_newly_merged_mixed_teacher",
    "unique_rows": {"original": 2000, "identity": 2000, "noisy": 2000, "total": 6000},
    "distillation": "completion_only_forward_kl",
    "temperature": 1.0,
    "fake_quantization": "Q4_0_STE",
    "optimizer": "Adafactor",
    "learning_rate": 0.000001,
    "weight_decay": 0.0,
    "effective_batch_size": 16,
    "optimizer_steps": 375,
    "product_quantization": "QAD-Q4_0"
  },
  "export": {
    "llama_cpp_tag": "b10158",
    "llama_cpp_revision": "f87067841bac583bc089a225382248d857791ca8",
    "tensor_contract": {"Q4_0": 92, "Q6_K": 1, "F32": 55, "total": 148}
  },
  "accepted_artifact": {
    "filename": "Scriber-LFM2.5-350M-Production-QAD-Q4_0.gguf",
    "size_bytes": 218328640,
    "sha256": "e1ca3391d896db64df91c5ed5a02e16f5b6bbec5de81667ec99535eb7b1c0486",
    "runtime_regression_exact": "600/600"
  },
  "attribution": "Praxist by Sapient Intelligence"
}