data/e4_gamma/EleutherAI--pythia-70m_random.json

{
  "model": "EleutherAI/pythia-70m",
  "theta": 10000,
  "gamma_pred": 0.99999561666838,
  "C_theory": 9.2103,
  "distances": [
    10,
    20,
    30,
    50,
    100,
    200,
    500,
    1000,
    2000
  ],
  "distances_fit": [
    30,
    50,
    100,
    200,
    500,
    1000,
    2000
  ],
  "d_min_fit": 30,
  "n_prompts_per_distance": 150,
  "seeds": [
    42,
    123,
    7
  ],
  "attn_by_distance": {
    "10": {
      "mean_across_seeds": 0.014303823270731502,
      "std_across_seeds": 0.0005389411970747021,
      "per_seed_means": [
        0.013846093355678022,
        0.01400491089404871,
        0.015060465562467774
      ]
    },
    "20": {
      "mean_across_seeds": 0.011754385376763014,
      "std_across_seeds": 0.0003750825584104145,
      "per_seed_means": [
        0.012279952564276754,
        0.011553768791879217,
        0.011429434774133067
      ]
    },
    "30": {
      "mean_across_seeds": 0.004743960113911372,
      "std_across_seeds": 0.0004785557585293483,
      "per_seed_means": [
        0.004255784423633789,
        0.004582108563821142,
        0.005393987354279185
      ]
    },
    "50": {
      "mean_across_seeds": 0.0035309766319632113,
      "std_across_seeds": 0.00030331972886044056,
      "per_seed_means": [
        0.003315780726649488,
        0.003317214461664359,
        0.003959934707575788
      ]
    },
    "100": {
      "mean_across_seeds": 0.0014786803761186699,
      "std_across_seeds": 4.0198997401330876e-05,
      "per_seed_means": [
        0.0015339826285101783,
        0.0014396193967938112,
        0.00146243910305202
      ]
    },
    "200": {
      "mean_across_seeds": 0.0007715668188918952,
      "std_across_seeds": 6.92475008380853e-05,
      "per_seed_means": [
        0.0008646439060976263,
        0.0007513971288184015,
        0.0006986594217596576
      ]
    },
    "500": {
      "mean_across_seeds": 0.0002215023043552517,
      "std_across_seeds": 2.714460980421038e-05,
      "per_seed_means": [
        0.00024024882055527995,
        0.00024114061049961797,
        0.0001831174820108572
      ]
    },
    "1000": {
      "mean_across_seeds": 8.175638578702798e-05,
      "std_across_seeds": 1.2207321420922821e-05,
      "per_seed_means": [
        8.717483436763965e-05,
        9.324252537529295e-05,
        6.485179761815137e-05
      ]
    },
    "2000": {
      "mean_across_seeds": 4.288268759245915e-05,
      "std_across_seeds": 9.477341124989812e-06,
      "per_seed_means": [
        3.2012329573566e-05,
        5.510804182828603e-05,
        4.152769137552544e-05
      ]
    }
  },
  "fit_power_law": {
    "gamma": 1.1705141984482668,
    "log_A": -1.1650521267460994,
    "R2": 0.994076,
    "n_points": 7
  },
  "fit_exponential": {
    "lambda": 0.00226006648331684,
    "log_A": -6.2565121323123964,
    "R2": 0.79868,
    "n_points": 7
  },
  "delta_R2_power_minus_exp": 0.1954,
  "gamma_CI_95": {
    "lo": 1.065560635977702,
    "hi": 1.268916819874023
  },
  "decision": "ANOMALY: long-context training effect"
}