taf-agent / data /e1_h3 /gpt2-xl_seed7.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": -6.1311,
"std": 2.8704
},
"baseline_logp_d1000": {
"mean": -10.6676,
"std": 1.8468
},
"logp_gap": 4.5365,
"layer_sweep": {
"0": {
"recovery_pct": -0.51,
"logp_transplant_mean": -10.6907,
"logp_transplant_std": 1.8202
},
"1": {
"recovery_pct": -1.59,
"logp_transplant_mean": -10.7395,
"logp_transplant_std": 1.8135
},
"2": {
"recovery_pct": -0.15,
"logp_transplant_mean": -10.6743,
"logp_transplant_std": 1.8851
},
"3": {
"recovery_pct": 3.68,
"logp_transplant_mean": -10.5005,
"logp_transplant_std": 2.0913
},
"4": {
"recovery_pct": 8.38,
"logp_transplant_mean": -10.2873,
"logp_transplant_std": 2.2274
},
"5": {
"recovery_pct": 5.83,
"logp_transplant_mean": -10.4031,
"logp_transplant_std": 2.1211
},
"6": {
"recovery_pct": 6.91,
"logp_transplant_mean": -10.3542,
"logp_transplant_std": 2.1677
},
"7": {
"recovery_pct": 5.58,
"logp_transplant_mean": -10.4145,
"logp_transplant_std": 2.1666
},
"8": {
"recovery_pct": 3.2,
"logp_transplant_mean": -10.5225,
"logp_transplant_std": 1.9956
},
"9": {
"recovery_pct": 2.34,
"logp_transplant_mean": -10.5616,
"logp_transplant_std": 1.9494
},
"10": {
"recovery_pct": 1.73,
"logp_transplant_mean": -10.5892,
"logp_transplant_std": 1.8673
},
"11": {
"recovery_pct": 2.45,
"logp_transplant_mean": -10.5565,
"logp_transplant_std": 1.8772
},
"12": {
"recovery_pct": -1.58,
"logp_transplant_mean": -10.7393,
"logp_transplant_std": 1.7629
},
"13": {
"recovery_pct": -1.05,
"logp_transplant_mean": -10.7154,
"logp_transplant_std": 1.769
},
"14": {
"recovery_pct": 2.11,
"logp_transplant_mean": -10.572,
"logp_transplant_std": 1.8271
},
"15": {
"recovery_pct": 4.28,
"logp_transplant_mean": -10.4733,
"logp_transplant_std": 1.88
},
"16": {
"recovery_pct": 9.44,
"logp_transplant_mean": -10.2393,
"logp_transplant_std": 2.0824
},
"17": {
"recovery_pct": 9.23,
"logp_transplant_mean": -10.2489,
"logp_transplant_std": 2.1382
},
"18": {
"recovery_pct": 9.06,
"logp_transplant_mean": -10.2567,
"logp_transplant_std": 2.1072
},
"19": {
"recovery_pct": 17.71,
"logp_transplant_mean": -9.8643,
"logp_transplant_std": 2.2154
},
"20": {
"recovery_pct": 22.84,
"logp_transplant_mean": -9.6317,
"logp_transplant_std": 2.3405
},
"21": {
"recovery_pct": 28.15,
"logp_transplant_mean": -9.3908,
"logp_transplant_std": 2.4645
},
"22": {
"recovery_pct": 34.35,
"logp_transplant_mean": -9.1092,
"logp_transplant_std": 2.5233
},
"23": {
"recovery_pct": 34.99,
"logp_transplant_mean": -9.0802,
"logp_transplant_std": 2.5107
},
"24": {
"recovery_pct": 35.89,
"logp_transplant_mean": -9.0393,
"logp_transplant_std": 2.4775
},
"25": {
"recovery_pct": 43.11,
"logp_transplant_mean": -8.712,
"logp_transplant_std": 2.6621
},
"26": {
"recovery_pct": 45.19,
"logp_transplant_mean": -8.6175,
"logp_transplant_std": 2.6992
},
"27": {
"recovery_pct": 44.75,
"logp_transplant_mean": -8.6374,
"logp_transplant_std": 2.6395
},
"28": {
"recovery_pct": 44.99,
"logp_transplant_mean": -8.6268,
"logp_transplant_std": 2.6259
},
"29": {
"recovery_pct": 49.18,
"logp_transplant_mean": -8.4365,
"logp_transplant_std": 2.7066
},
"30": {
"recovery_pct": 49.9,
"logp_transplant_mean": -8.404,
"logp_transplant_std": 2.6381
},
"31": {
"recovery_pct": 48.1,
"logp_transplant_mean": -8.4854,
"logp_transplant_std": 2.5323
},
"32": {
"recovery_pct": 52.8,
"logp_transplant_mean": -8.2725,
"logp_transplant_std": 2.6267
},
"33": {
"recovery_pct": 53.24,
"logp_transplant_mean": -8.2523,
"logp_transplant_std": 2.5726
},
"34": {
"recovery_pct": 56.57,
"logp_transplant_mean": -8.1011,
"logp_transplant_std": 2.5822
},
"35": {
"recovery_pct": 62.14,
"logp_transplant_mean": -7.8488,
"logp_transplant_std": 2.6612
},
"36": {
"recovery_pct": 67.02,
"logp_transplant_mean": -7.627,
"logp_transplant_std": 2.687
},
"37": {
"recovery_pct": 68.45,
"logp_transplant_mean": -7.5624,
"logp_transplant_std": 2.746
},
"38": {
"recovery_pct": 73.3,
"logp_transplant_mean": -7.3424,
"logp_transplant_std": 2.7967
},
"39": {
"recovery_pct": 78.31,
"logp_transplant_mean": -7.115,
"logp_transplant_std": 2.8228
},
"40": {
"recovery_pct": 81.57,
"logp_transplant_mean": -6.9672,
"logp_transplant_std": 2.786
},
"41": {
"recovery_pct": 86.57,
"logp_transplant_mean": -6.7404,
"logp_transplant_std": 2.8
},
"42": {
"recovery_pct": 89.61,
"logp_transplant_mean": -6.6023,
"logp_transplant_std": 2.8288
},
"43": {
"recovery_pct": 93.05,
"logp_transplant_mean": -6.4463,
"logp_transplant_std": 2.8746
},
"44": {
"recovery_pct": 96.06,
"logp_transplant_mean": -6.3099,
"logp_transplant_std": 2.8621
},
"45": {
"recovery_pct": 98.67,
"logp_transplant_mean": -6.1914,
"logp_transplant_std": 2.8952
},
"46": {
"recovery_pct": 100.17,
"logp_transplant_mean": -6.1233,
"logp_transplant_std": 2.8761
},
"47": {
"recovery_pct": 100.54,
"logp_transplant_mean": -6.1065,
"logp_transplant_std": 2.8737
}
},
"L_crit_90": 43,
"L_crit_99": 46,
"alpha_90": 0.8958,
"alpha_99": 0.9583,
"recovery_at_Lcrit": 0.9305,
"n_layers": 48,
"d_model": 1600,
"model": "gpt2-xl",
"seed": 7,
"runtime_seconds": 1254.7
}