{
  "title": "LoRA Training History",
  "latest_checkpoint_id": "v0007__20260319T2126Z",
  "latest_run_id": "unsloth-sft-1773952401",
  "saved_checkpoint_count": 7,
  "base_model_name": "unsloth/gpt-oss-20b-unsloth-bnb-4bit",
  "peft_type": "LORA",
  "lora_rank": 8,
  "lora_alpha": 16,
  "lora_dropout": 0,
  "target_modules_count": 60,
  "instance_types": [
    "ml.g6e.xlarge"
  ],
  "regions": [
    "us-east-2"
  ],
  "sft_modes": [
    "llm_io"
  ],
  "max_token_sequence_length": [
    4600
  ],
  "sequence_constraint": [
    4600
  ],
  "infer_max_completion": [
    600
  ],
  "max_llm_io_examples": [
    168
  ],
  "sft_pairs_line_count": 951,
  "code_revisions": [
    "5f8220f87ed5",
    "8fda0e9d7f87",
    "95e48bbf1c35",
    "b536711d6a19",
    "6fd51d77784a",
    "cf8fc126d7be",
    "eb55d49aa12f"
  ],
  "checkpoint_date_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"
  },
  "latest_run_dataset_examples": 168,
  "latest_run_observation_count": 30,
  "latest_run_definition_count": 1824,
  "latest_archive_successful_non_replace_observation_count": 32,
  "package_versions": {
    "torch": "2.9.1+cu128",
    "transformers": "4.57.1",
    "peft": "0.18.0"
  },
  "training_stack": [
    "UnSloth",
    "PyTorch 2.9.1+cu128",
    "Transformers 4.57.1",
    "PEFT 0.18.0"
  ],
  "training_software": "UnSloth",
  "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",
  "latest_training_instance_type": "ml.g6e.xlarge",
  "latest_training_region": "us-east-2",
  "latest_sft_mode": "llm_io",
  "latest_sequence_length": 4600,
  "latest_infer_max_completion": 600
}