{
  "source_notebook": "Advanced_Llama3_2_(3B)_GRPO_LoRA.ipynb",
  "source_sha256": "a9a33abbb0f691ab5405bfa1b8b1b1cbd5e6aa21c2534bf76e5879766dbed6fa",
  "record_type": "Retained notebook outputs, not a newly reproduced GPU run",
  "training_steps": 500,
  "trainer_runtime_seconds": 25114.7034,
  "trainable_parameters": 97255424,
  "reported_total_parameters": 3310005248,
  "mean_logged_reward": {
    "steps_1_to_50": -0.559375,
    "steps_451_to_500": 7.4925
  },
  "first_logged_step": {
    "Step": "1",
    "Training Loss": "0.000000",
    "reward": "-0.656250",
    "reward_std": "0.780387",
    "completion_length": "242.437500",
    "kl": "0.000000",
    "rewards / match_format_exactly": "0.000000",
    "rewards / match_format_approximately": "-1.093750",
    "rewards / check_answer": "0.000000",
    "rewards / check_numbers": "0.437500"
  },
  "last_logged_step": {
    "Step": "500",
    "Training Loss": "0.005800",
    "reward": "9.218750",
    "reward_std": "0.562500",
    "completion_length": "145.625000",
    "kl": "0.145031",
    "rewards / match_format_exactly": "3.000000",
    "rewards / match_format_approximately": "2.000000",
    "rewards / check_answer": "2.718750",
    "rewards / check_numbers": "1.500000"
  },
  "caveats": [
    "Rewards combine format and answer checks; they are not held-out accuracy.",
    "Visible cells and installation versions are not a complete immutable record of the saved run.",
    "The baseline and adapter inference examples use different system prompting."
  ]
}