taf-agent / data /e1_h3 /gpt2-xl_seed42.json
karlexmarin's picture
feat: ship paper artefacts + CLI diagnostic alongside browser tool
535348a
raw
history blame
6.43 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": -6.0101,
"std": 2.7684
},
"baseline_logp_d1000": {
"mean": -10.7722,
"std": 1.8118
},
"logp_gap": 4.762,
"layer_sweep": {
"0": {
"recovery_pct": -0.67,
"logp_transplant_mean": -10.8041,
"logp_transplant_std": 1.7484
},
"1": {
"recovery_pct": -1.6,
"logp_transplant_mean": -10.8484,
"logp_transplant_std": 1.7243
},
"2": {
"recovery_pct": -0.62,
"logp_transplant_mean": -10.8014,
"logp_transplant_std": 1.7603
},
"3": {
"recovery_pct": 1.54,
"logp_transplant_mean": -10.6988,
"logp_transplant_std": 1.7851
},
"4": {
"recovery_pct": 5.86,
"logp_transplant_mean": -10.4929,
"logp_transplant_std": 1.9665
},
"5": {
"recovery_pct": 3.69,
"logp_transplant_mean": -10.5964,
"logp_transplant_std": 1.8037
},
"6": {
"recovery_pct": 6.1,
"logp_transplant_mean": -10.4815,
"logp_transplant_std": 1.8284
},
"7": {
"recovery_pct": 5.72,
"logp_transplant_mean": -10.4996,
"logp_transplant_std": 1.8839
},
"8": {
"recovery_pct": 4.45,
"logp_transplant_mean": -10.5603,
"logp_transplant_std": 1.769
},
"9": {
"recovery_pct": 3.62,
"logp_transplant_mean": -10.5999,
"logp_transplant_std": 1.7496
},
"10": {
"recovery_pct": 3.29,
"logp_transplant_mean": -10.6155,
"logp_transplant_std": 1.6963
},
"11": {
"recovery_pct": 4.07,
"logp_transplant_mean": -10.5785,
"logp_transplant_std": 1.6989
},
"12": {
"recovery_pct": 2.17,
"logp_transplant_mean": -10.6686,
"logp_transplant_std": 1.6391
},
"13": {
"recovery_pct": 2.84,
"logp_transplant_mean": -10.637,
"logp_transplant_std": 1.6874
},
"14": {
"recovery_pct": 4.28,
"logp_transplant_mean": -10.5683,
"logp_transplant_std": 1.7259
},
"15": {
"recovery_pct": 5.38,
"logp_transplant_mean": -10.5161,
"logp_transplant_std": 1.7665
},
"16": {
"recovery_pct": 9.68,
"logp_transplant_mean": -10.3111,
"logp_transplant_std": 1.9126
},
"17": {
"recovery_pct": 9.41,
"logp_transplant_mean": -10.3239,
"logp_transplant_std": 1.9505
},
"18": {
"recovery_pct": 8.21,
"logp_transplant_mean": -10.3814,
"logp_transplant_std": 1.9049
},
"19": {
"recovery_pct": 14.67,
"logp_transplant_mean": -10.0734,
"logp_transplant_std": 1.9501
},
"20": {
"recovery_pct": 21.1,
"logp_transplant_mean": -9.7676,
"logp_transplant_std": 2.1252
},
"21": {
"recovery_pct": 27.27,
"logp_transplant_mean": -9.4733,
"logp_transplant_std": 2.2339
},
"22": {
"recovery_pct": 34.0,
"logp_transplant_mean": -9.1532,
"logp_transplant_std": 2.3275
},
"23": {
"recovery_pct": 36.19,
"logp_transplant_mean": -9.0486,
"logp_transplant_std": 2.382
},
"24": {
"recovery_pct": 37.71,
"logp_transplant_mean": -8.9764,
"logp_transplant_std": 2.3272
},
"25": {
"recovery_pct": 45.5,
"logp_transplant_mean": -8.6056,
"logp_transplant_std": 2.5183
},
"26": {
"recovery_pct": 46.71,
"logp_transplant_mean": -8.5477,
"logp_transplant_std": 2.5544
},
"27": {
"recovery_pct": 45.42,
"logp_transplant_mean": -8.6094,
"logp_transplant_std": 2.464
},
"28": {
"recovery_pct": 45.21,
"logp_transplant_mean": -8.6192,
"logp_transplant_std": 2.4324
},
"29": {
"recovery_pct": 47.78,
"logp_transplant_mean": -8.4967,
"logp_transplant_std": 2.4581
},
"30": {
"recovery_pct": 47.68,
"logp_transplant_mean": -8.5017,
"logp_transplant_std": 2.3658
},
"31": {
"recovery_pct": 46.1,
"logp_transplant_mean": -8.5767,
"logp_transplant_std": 2.2636
},
"32": {
"recovery_pct": 50.55,
"logp_transplant_mean": -8.3651,
"logp_transplant_std": 2.3411
},
"33": {
"recovery_pct": 52.65,
"logp_transplant_mean": -8.2651,
"logp_transplant_std": 2.3238
},
"34": {
"recovery_pct": 57.13,
"logp_transplant_mean": -8.0517,
"logp_transplant_std": 2.3993
},
"35": {
"recovery_pct": 61.2,
"logp_transplant_mean": -7.8579,
"logp_transplant_std": 2.4713
},
"36": {
"recovery_pct": 66.57,
"logp_transplant_mean": -7.602,
"logp_transplant_std": 2.5163
},
"37": {
"recovery_pct": 68.86,
"logp_transplant_mean": -7.4932,
"logp_transplant_std": 2.544
},
"38": {
"recovery_pct": 72.04,
"logp_transplant_mean": -7.3417,
"logp_transplant_std": 2.5854
},
"39": {
"recovery_pct": 77.66,
"logp_transplant_mean": -7.0738,
"logp_transplant_std": 2.66
},
"40": {
"recovery_pct": 79.9,
"logp_transplant_mean": -6.9673,
"logp_transplant_std": 2.6734
},
"41": {
"recovery_pct": 84.68,
"logp_transplant_mean": -6.7399,
"logp_transplant_std": 2.7111
},
"42": {
"recovery_pct": 87.7,
"logp_transplant_mean": -6.596,
"logp_transplant_std": 2.7066
},
"43": {
"recovery_pct": 91.05,
"logp_transplant_mean": -6.4362,
"logp_transplant_std": 2.7429
},
"44": {
"recovery_pct": 94.6,
"logp_transplant_mean": -6.2671,
"logp_transplant_std": 2.7547
},
"45": {
"recovery_pct": 98.51,
"logp_transplant_mean": -6.0812,
"logp_transplant_std": 2.799
},
"46": {
"recovery_pct": 99.8,
"logp_transplant_mean": -6.0198,
"logp_transplant_std": 2.8075
},
"47": {
"recovery_pct": 100.52,
"logp_transplant_mean": -5.9854,
"logp_transplant_std": 2.8109
}
},
"L_crit_90": 43,
"L_crit_99": 46,
"alpha_90": 0.8958,
"alpha_99": 0.9583,
"recovery_at_Lcrit": 0.9105,
"n_layers": 48,
"d_model": 1600,
"model": "gpt2-xl",
"seed": 42,
"runtime_seconds": 1237.8
}