{
  "schema_version": 2,
  "date": "2026-08-20",
  "scope": "engineering_smoke_only",
  "seeds": [
    42
  ],
  "hardware": {
    "device": "cuda",
    "gpu": "NVIDIA A30 24GB",
    "mixed_precision": "bf16"
  },
  "model": {
    "id": "HuggingFaceTB/SmolLM2-135M-Instruct",
    "revision": "12fd25f77366fa6b3b4b768ec3050bf629380bac"
  },
  "reasoning_evolve": {
    "dataset": "arithmetic-generate",
    "evaluation_examples": 2,
    "baseline": {
      "samples": 1,
      "accuracy": 0.0,
      "generated_tokens": 0.0,
      "latency_seconds": 0.43901340290904045
    },
    "candidate": {
      "samples": 2,
      "accuracy": 0.0,
      "generated_tokens": 152.0,
      "tokens_per_example": 76.0,
      "latency_seconds": 3.298396397382021
    },
    "result": "completed_with_markdown_and_json_report"
  },
  "gsm8k_sft": {
    "train_examples": 8,
    "batch_size": 2,
    "step_1_loss": 1.5169668197631836,
    "resumed_step_2_loss": 1.3653311729431152,
    "resumed_from_step": 1,
    "three_seed_accuracy": 0.0,
    "result": "parameter_update_checkpoint_and_optimizer_resume_completed"
  },
  "ultrafeedback": {
    "dataset_id": "HuggingFaceH4/ultrafeedback_binarized",
    "revision": "292c16329d921287c4166934cac1a6ad1e13a6c5",
    "train_examples": 4,
    "heldout_examples": 2,
    "dpo": {
      "steps": 1,
      "loss": 0.6931471824645996,
      "heldout_preference_accuracy": 0.0
    },
    "orpo": {
      "steps": 1,
      "loss": 1.3450264930725098,
      "heldout_preference_accuracy": 0.0
    },
    "result": "both_real_parameter_update_paths_completed"
  },
  "limitations": [
    "The reduced examples and steps validate engineering integrity only.",
    "Zero accuracy is reported as observed and is not an algorithm conclusion.",
    "Checkpoints and dataset samples remain outside Git."
  ],
  "manifest_ref": "experiments:a30-reasoning-posttraining-smoke-20260820",
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single/few-seed reduced-step engineering smoke; do not claim a stable improvement"
  },
  "provenance": {
    "artifact_path": "docs/experiments/a30-reasoning-posttraining-smoke-20260820.json",
    "dataset_fingerprint": {
      "gsm8k_train_sha256": "17f347dc51477c50d4efb83959dbb7c56297aba886e5544ee2aaed3024813465",
      "gsm8k_test_sha256": "3730d312f6e3440559ace48831e51066acaca737f6eabec99bccb9e4b3c39d14",
      "ultrafeedback_train_subset_sha256": "d273be5a5f0ee33bc6b362db59d8973398e590d2d2e82c85fe65827372e538ed",
      "ultrafeedback_test_subset_sha256": "8cce8a46b00e29a2376310c9be433bd5508a347aa48a06fa9717b655f0c6f385"
    }
  }
}