{
  "package_slug": "lora_training_history",
  "title": "LoRA Training History",
  "summary": "Saved adapter-checkpoint lineage, training configuration, dataset growth, stability diagnostics, and reward-consumption history for the current fine-tuned model runs.",
  "overview": {
    "saved_checkpoint_count": 7,
    "latest_checkpoint_id": "v0007__20260319T2126Z",
    "latest_run_id": "unsloth-sft-1773952401",
    "checkpoint_span": {
      "first_checkpoint_created_at": "2026-03-03T03:02:53Z",
      "latest_checkpoint_created_at": "2026-03-19T21:26:41Z",
      "first_checkpoint_date_long": "March 3, 2026",
      "latest_checkpoint_date_long": "March 19, 2026"
    },
    "base_model_name": "unsloth/gpt-oss-20b-unsloth-bnb-4bit",
    "training_stack": [
      "UnSloth",
      "PyTorch 2.9.1+cu128",
      "Transformers 4.57.1",
      "PEFT 0.18.0"
    ],
    "fine_tuning_method": "Reward-weighted SFT on credited llm_output artifacts from successful observations",
    "serving_setup": "Base model plus the latest LoRA adapter checkpoint",
    "training_stabilizer": "clip_higher",
    "training_stabilizer_explanation": "clip_higher ratio clipping against cached completion log-probabilities to limit destabilizing policy jumps",
    "training_precision": "bf16",
    "optimizer": "adamw_8bit",
    "learning_rates": [
      "4e-05",
      "5e-05"
    ],
    "hyperparameter_tuning_note": "We tuned some training-job parameters across runs while keeping the core setup fixed. Learning rate was 5e-05 through v6 and changed to 4e-05 at v7. A 168-example training cap was introduced at v4. Core model, LoRA shape, optimizer, precision, and hardware remained the same.",
    "checkpoint_progression": "Each saved adapter resumes from the previous saved adapter when available",
    "max_token_sequence_length": [
      4600
    ],
    "lora_rank": 8,
    "lora_alpha": 16,
    "target_modules_count": 60,
    "latest_run_dataset_examples": 168,
    "latest_run_observation_count": 30,
    "latest_run_definition_count": 1824,
    "latest_archive_successful_non_replace_observation_count": 32,
    "latest_training_instance_type": "ml.g6e.xlarge",
    "latest_training_region": "us-east-2",
    "latest_sft_mode": "llm_io",
    "sft_pairs_line_count": 951
  },
  "files": {
    "README.md": "Plain-language overview of the LoRA training-history artifact.",
    "summaries/checkpoint_history.csv": "One row per saved checkpoint with configuration, dataset scale, and top-line metrics.",
    "summaries/training_setup_summary.json": "Stable training setup summary, including LoRA parameters and max token sequence length.",
    "summaries/dataset_growth_summary.json": "Per-version growth in examples, rewarded observations, and tracked definitions.",
    "summaries/reward_consumption_summary.csv": "Per-checkpoint reward consumed by that training job and reward still pending in the archive afterward.",
    "summaries/reward_consumption_summary.json": "JSON form of the consumed-versus-pending reward history across all saved versions.",
    "figures/dataset_growth.png": "Examples, rewarded observations, and definition counts across saved versions.",
    "figures/training_stability_by_checkpoint.png": "Checkpoint reward scalar and clipped-token share across saved versions.",
    "figures/reward_consumption_by_checkpoint.png": "Reward consumed by each training job versus reward still pending in the archive afterward.",
    "source_records/latest_checkpoint_provenance.json": "Public latest-checkpoint provenance summary with code revision, package versions, adapter shape, and dataset counts.",
    "source_records/latest_reward_consumption_receipt.json": "Latest reward-consumption receipt showing what reward was applied at the newest checkpoint.",
    "source_records/latest_run_record.json": "Public latest-run record with trainer scalar summaries and distributions."
  }
}