{
  "schema_version": 2,
  "kind": "empirical_micro_lm_scaling_curve",
  "date": "2026-08-20",
  "seed": 42,
  "scope": "four-point engineering smoke on local Mac CPU with an A30 CUDA replay",
  "fixed_axes": {
    "dataset": "WikiText-2",
    "architecture": "gpt_baseline",
    "vocab_size": 1024,
    "batch_size": 2,
    "sequence_length": 64,
    "optimizer": "adamw",
    "learning_rate": 0.0003
  },
  "points": [
    {
      "id": "scale-0",
      "dimensions": 64,
      "layers": 2,
      "available_train_tokens": 12000,
      "steps": 6,
      "parameters": 169728,
      "tokens_seen": 768,
      "estimated_training_flops": 782106624,
      "validation_lm_loss": 6.867805019021034,
      "predicted_validation_lm_loss": 6.879698965177925,
      "relative_fit_residual": -0.0017318557467200827
    },
    {
      "id": "scale-1",
      "dimensions": 64,
      "layers": 2,
      "available_train_tokens": 24000,
      "steps": 12,
      "parameters": 169728,
      "tokens_seen": 1536,
      "estimated_training_flops": 1564213248,
      "validation_lm_loss": 6.798588290810585,
      "predicted_validation_lm_loss": 6.797736379500577,
      "relative_fit_residual": 0.00012530610346126844
    },
    {
      "id": "scale-2",
      "dimensions": 96,
      "layers": 2,
      "available_train_tokens": 24000,
      "steps": 12,
      "parameters": 328320,
      "tokens_seen": 1536,
      "estimated_training_flops": 3025797120,
      "validation_lm_loss": 6.743992105126381,
      "predicted_validation_lm_loss": 6.720624606391065,
      "relative_fit_residual": 0.0034648815425466807
    },
    {
      "id": "scale-3",
      "dimensions": 128,
      "layers": 3,
      "available_train_tokens": 48000,
      "steps": 18,
      "parameters": 734336,
      "tokens_seen": 2304,
      "estimated_training_flops": 10151460864,
      "validation_lm_loss": 6.569156587123871,
      "predicted_validation_lm_loss": 6.581424610277404,
      "relative_fit_residual": -0.0018675403744240937
    }
  ],
  "fit": {
    "compute_power_law": {
      "log_compute_slope": -0.017291103277517088,
      "descriptive_alpha": 0.017291103277517088,
      "log_rmse": 0.0021528097797185236,
      "loss_rmse": 0.014480446405229906,
      "r_squared": 0.9829283084779702,
      "max_absolute_relative_error": 0.0034648815425466807
    },
    "parameter_data_surface": {
      "status": "descriptive_only",
      "log_parameter_slope": -0.015559764148433915,
      "log_data_slope": -0.015629893675293778,
      "log_rmse": 0.0011925184971653651,
      "loss_rmse": 0.00802999249256376,
      "r_squared": 0.9947616382382629,
      "max_absolute_relative_error": 0.002002566182144204
    }
  },
  "a30_cuda_replay": {
    "validation_lm_losses": [
      6.867805063724518,
      6.798588261008263,
      6.743992269039154,
      6.569156527519226
    ],
    "log_compute_slope": -0.017291107035388338,
    "log_rmse": 0.0021528201704139804,
    "r_squared": 0.9829281537891447,
    "all_points_resolved_device": "cuda"
  },
  "claim_boundary": "Local descriptive reduced-step fit only; not a Chinchilla compute-optimal frontier or a large-model extrapolation.",
  "manifest_ref": "experiments:fm002-scaling-law-mac-seed42",
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single-seed multi-budget engineering smoke; do not claim a stable scaling exponent"
  },
  "provenance": {
    "artifact_path": "docs/experiments/fm002-scaling-law-mac-seed42.json",
    "dataset_fingerprint": {
      "wikitext2_train_sha256": "9e9fa1ad55b1c2c95b08e37dd8e653f638fac2c6de904b79e813611eefbc985f",
      "wikitext2_validation_sha256": "f0737ed31fc1329026e95cb8b98e19c2a182c39c240ab909dc31abf2f8af58e8",
      "wikitext2_test_sha256": "d790b833ef8cf03a90db7bf1271b7520b83c45ce07ba3c1a9699df81e239eca0",
      "tokenizer_sha256": "348bfd9afd61d14a6966069f1d37d8ee238ba53d9adb0d62c5745c4a87763e25"
    }
  }
}