taf-agent / data /e1_h3 /gpt2-xl_seed123.json
karlexmarin's picture
feat: ship paper artefacts + CLI diagnostic alongside browser tool
535348a
raw
history blame
6.44 kB
{
"tau_90": 0.9,
"tau_99": 0.99,
"N_prompts": 50,
"dist_short": 10,
"dist_long": 1000,
"vocab_low": 1000,
"vocab_high": 49000,
"n_prefix": 20,
"ref_logp_d10": {
"mean": -6.1068,
"std": 2.589
},
"baseline_logp_d1000": {
"mean": -10.7776,
"std": 1.6556
},
"logp_gap": 4.6708,
"layer_sweep": {
"0": {
"recovery_pct": -0.74,
"logp_transplant_mean": -10.812,
"logp_transplant_std": 1.596
},
"1": {
"recovery_pct": -2.23,
"logp_transplant_mean": -10.8816,
"logp_transplant_std": 1.5664
},
"2": {
"recovery_pct": -1.81,
"logp_transplant_mean": -10.8623,
"logp_transplant_std": 1.5771
},
"3": {
"recovery_pct": 0.6,
"logp_transplant_mean": -10.7495,
"logp_transplant_std": 1.7063
},
"4": {
"recovery_pct": 8.48,
"logp_transplant_mean": -10.3813,
"logp_transplant_std": 2.002
},
"5": {
"recovery_pct": 4.55,
"logp_transplant_mean": -10.565,
"logp_transplant_std": 1.8072
},
"6": {
"recovery_pct": 6.52,
"logp_transplant_mean": -10.4731,
"logp_transplant_std": 1.8261
},
"7": {
"recovery_pct": 6.53,
"logp_transplant_mean": -10.4727,
"logp_transplant_std": 1.8417
},
"8": {
"recovery_pct": 2.93,
"logp_transplant_mean": -10.6406,
"logp_transplant_std": 1.7092
},
"9": {
"recovery_pct": 2.08,
"logp_transplant_mean": -10.6803,
"logp_transplant_std": 1.6812
},
"10": {
"recovery_pct": 1.87,
"logp_transplant_mean": -10.6901,
"logp_transplant_std": 1.6302
},
"11": {
"recovery_pct": 2.56,
"logp_transplant_mean": -10.6581,
"logp_transplant_std": 1.6429
},
"12": {
"recovery_pct": -1.15,
"logp_transplant_mean": -10.8315,
"logp_transplant_std": 1.5511
},
"13": {
"recovery_pct": -0.81,
"logp_transplant_mean": -10.8157,
"logp_transplant_std": 1.5128
},
"14": {
"recovery_pct": 1.63,
"logp_transplant_mean": -10.7017,
"logp_transplant_std": 1.5034
},
"15": {
"recovery_pct": 3.61,
"logp_transplant_mean": -10.6091,
"logp_transplant_std": 1.4998
},
"16": {
"recovery_pct": 8.83,
"logp_transplant_mean": -10.3652,
"logp_transplant_std": 1.589
},
"17": {
"recovery_pct": 8.39,
"logp_transplant_mean": -10.386,
"logp_transplant_std": 1.6282
},
"18": {
"recovery_pct": 7.6,
"logp_transplant_mean": -10.4229,
"logp_transplant_std": 1.5825
},
"19": {
"recovery_pct": 15.19,
"logp_transplant_mean": -10.0684,
"logp_transplant_std": 1.6868
},
"20": {
"recovery_pct": 21.97,
"logp_transplant_mean": -9.7514,
"logp_transplant_std": 1.7709
},
"21": {
"recovery_pct": 29.16,
"logp_transplant_mean": -9.4156,
"logp_transplant_std": 1.8673
},
"22": {
"recovery_pct": 36.89,
"logp_transplant_mean": -9.0546,
"logp_transplant_std": 1.9927
},
"23": {
"recovery_pct": 39.04,
"logp_transplant_mean": -8.9541,
"logp_transplant_std": 2.0406
},
"24": {
"recovery_pct": 40.02,
"logp_transplant_mean": -8.9082,
"logp_transplant_std": 2.009
},
"25": {
"recovery_pct": 47.9,
"logp_transplant_mean": -8.5403,
"logp_transplant_std": 2.1029
},
"26": {
"recovery_pct": 50.3,
"logp_transplant_mean": -8.4283,
"logp_transplant_std": 2.1486
},
"27": {
"recovery_pct": 49.85,
"logp_transplant_mean": -8.449,
"logp_transplant_std": 2.113
},
"28": {
"recovery_pct": 49.81,
"logp_transplant_mean": -8.4513,
"logp_transplant_std": 2.1829
},
"29": {
"recovery_pct": 53.53,
"logp_transplant_mean": -8.2772,
"logp_transplant_std": 2.2809
},
"30": {
"recovery_pct": 54.81,
"logp_transplant_mean": -8.2176,
"logp_transplant_std": 2.1936
},
"31": {
"recovery_pct": 52.54,
"logp_transplant_mean": -8.3237,
"logp_transplant_std": 2.0964
},
"32": {
"recovery_pct": 56.03,
"logp_transplant_mean": -8.1607,
"logp_transplant_std": 2.1807
},
"33": {
"recovery_pct": 56.67,
"logp_transplant_mean": -8.1305,
"logp_transplant_std": 2.1518
},
"34": {
"recovery_pct": 61.18,
"logp_transplant_mean": -7.9201,
"logp_transplant_std": 2.2444
},
"35": {
"recovery_pct": 64.91,
"logp_transplant_mean": -7.7458,
"logp_transplant_std": 2.2279
},
"36": {
"recovery_pct": 69.54,
"logp_transplant_mean": -7.5297,
"logp_transplant_std": 2.2359
},
"37": {
"recovery_pct": 71.35,
"logp_transplant_mean": -7.4449,
"logp_transplant_std": 2.2532
},
"38": {
"recovery_pct": 74.03,
"logp_transplant_mean": -7.3196,
"logp_transplant_std": 2.2638
},
"39": {
"recovery_pct": 78.54,
"logp_transplant_mean": -7.1094,
"logp_transplant_std": 2.3269
},
"40": {
"recovery_pct": 81.49,
"logp_transplant_mean": -6.9715,
"logp_transplant_std": 2.3364
},
"41": {
"recovery_pct": 86.23,
"logp_transplant_mean": -6.7502,
"logp_transplant_std": 2.3837
},
"42": {
"recovery_pct": 88.99,
"logp_transplant_mean": -6.6209,
"logp_transplant_std": 2.4614
},
"43": {
"recovery_pct": 92.43,
"logp_transplant_mean": -6.4604,
"logp_transplant_std": 2.5157
},
"44": {
"recovery_pct": 95.49,
"logp_transplant_mean": -6.3174,
"logp_transplant_std": 2.5227
},
"45": {
"recovery_pct": 98.42,
"logp_transplant_mean": -6.1808,
"logp_transplant_std": 2.5885
},
"46": {
"recovery_pct": 99.97,
"logp_transplant_mean": -6.1084,
"logp_transplant_std": 2.5872
},
"47": {
"recovery_pct": 100.54,
"logp_transplant_mean": -6.0817,
"logp_transplant_std": 2.6
}
},
"L_crit_90": 43,
"L_crit_99": 46,
"alpha_90": 0.8958,
"alpha_99": 0.9583,
"recovery_at_Lcrit": 0.9243,
"n_layers": 48,
"d_model": 1600,
"model": "gpt2-xl",
"seed": 123,
"runtime_seconds": 1241.2
}