data/checkpoint_eval/pythia-70m_checkpoint_comparison.json

{
  "step8000": {
    "revision": "step8000",
    "expected_gamma": 1.064,
    "gamma_obs": 1.0266835966009398,
    "r2": 0.9820928579517015,
    "attn_means_per_distance": [
      0.013771302501360575,
      0.008232926908466548,
      0.00256258381737603,
      0.002347815906008085,
      0.0010662785420815148,
      0.0002845022413465712,
      0.00010495591494772167,
      7.333743075529732e-05
    ],
    "passkey_accuracy": {
      "256": 0.05,
      "512": 0.2,
      "1024": 0.05,
      "2048": 0.0
    },
    "note_public_benchmarks": "LAMBADA/HellaSwag/PPL available in combined_training_dynamics.json from EleutherAI public evals"
  },
  "step143000": {
    "revision": "step143000",
    "expected_gamma": 0.781,
    "gamma_obs": NaN,
    "r2": NaN,
    "attn_means_per_distance": [
      NaN,
      NaN,
      NaN,
      NaN,
      NaN,
      NaN,
      NaN,
      NaN
    ],
    "passkey_accuracy": {
      "256": 0.0,
      "512": 0.0,
      "1024": 0.0,
      "2048": 0.0
    },
    "note_public_benchmarks": "LAMBADA/HellaSwag/PPL available in combined_training_dynamics.json from EleutherAI public evals"
  }
}