arch_findings.json

{
  "n_models": 4,
  "n_with_macro": 4,
  "correlations": {
    "total_params": {
      "r": -0.13171951916328523,
      "n": 4
    },
    "n_layers": {
      "r": -0.5375124456001662,
      "n": 4
    },
    "d_model": {
      "r": 0.5671372674429019,
      "n": 4
    },
    "n_heads": {
      "r": 0.5671372674429019,
      "n": 4
    },
    "ffn_dim": {
      "r": 0.9208998623145593,
      "n": 3
    },
    "vocab_size": {
      "r": 0.3646043405335206,
      "n": 4
    },
    "max_ctx": {
      "r": 0.1538621267359702,
      "n": 4
    }
  },
  "correlations_ranked": [
    [
      "ffn_dim",
      0.9208998623145593,
      3
    ],
    [
      "d_model",
      0.5671372674429019,
      4
    ],
    [
      "n_heads",
      0.5671372674429019,
      4
    ],
    [
      "n_layers",
      -0.5375124456001662,
      4
    ],
    [
      "vocab_size",
      0.3646043405335206,
      4
    ],
    [
      "max_ctx",
      0.1538621267359702,
      4
    ],
    [
      "total_params",
      -0.13171951916328523,
      4
    ]
  ],
  "best_macro_model": {
    "repo_id": "exnivo/tinybrain-100m-base",
    "macro": 0.5120183232855188,
    "arch": {
      "total_params": 103385856,
      "n_layers": 12,
      "d_model": 768,
      "n_heads": 12,
      "ffn_dim": 2048,
      "vocab_size": 24000,
      "max_ctx": 2048
    }
  },
  "models": [
    {
      "repo_id": "exnivo/tinybrain-100m-base",
      "macro": 0.5120183232855188,
      "total_params": 103385856,
      "model_type": "llama",
      "n_layers": 12,
      "d_model": 768
    },
    {
      "repo_id": "aksern/nexi-g1",
      "macro": 0.4822002572056585,
      "total_params": 30339456,
      "model_type": "gpt2",
      "n_layers": 6,
      "d_model": 384
    },
    {
      "repo_id": "oddadmix/Emhotob-25M-Egyptian-English-v2",
      "macro": 0.39475337003755284,
      "total_params": 25271424,
      "model_type": "llama",
      "n_layers": 8,
      "d_model": 384
    },
    {
      "repo_id": "textilelabs/Loom-Crucible-Preview",
      "macro": 0.391990733057559,
      "total_params": 154980864,
      "model_type": "llama",
      "n_layers": 52,
      "d_model": 512
    }
  ],
  "caveats": [
    "n is very small (4 models); correlations are illustrative, not statistical.",
    "All scores are zero-shot loglikelihood on a single harness (lm-eval 0.4.13, float32, bs=8, cuda:0).",
    "BLiMP is the mean acc over its subtasks; ARC-Easy/PIQA are acc; HellaSwag is acc_norm.",
    "Models span different training corpora and token counts, so arch and data effects are confounded.",
    "A tiny model trained on a narrow domain can score well on one task and poorly on another; macro hides that."
  ]
}