taf-agent / data /e1_test /EleutherAI--pythia-1b_seed7.json
karlexmarin's picture
feat: ship paper artefacts + CLI diagnostic alongside browser tool
535348a
raw
history blame
2.53 kB
{
"tau_90": 0.9,
"tau_99": 0.99,
"N_prompts": 2,
"dist_short": 10,
"dist_long": 2048,
"vocab_low": 1000,
"vocab_high": 49000,
"n_prefix": 20,
"ref_logp_d10": {
"mean": -9.2085,
"std": 1.2153
},
"baseline_logp_d1000": {
"mean": -12.8734,
"std": 0.1521
},
"logp_gap": 3.6649,
"layer_sweep": {
"0": {
"recovery_pct": 0.0,
"logp_transplant_mean": -12.8734,
"logp_transplant_std": 0.1521
},
"1": {
"recovery_pct": -1.05,
"logp_transplant_mean": -12.9117,
"logp_transplant_std": 0.078
},
"2": {
"recovery_pct": -3.08,
"logp_transplant_mean": -12.9862,
"logp_transplant_std": 0.0051
},
"3": {
"recovery_pct": -1.04,
"logp_transplant_mean": -12.9115,
"logp_transplant_std": 0.0218
},
"4": {
"recovery_pct": -2.51,
"logp_transplant_mean": -12.9655,
"logp_transplant_std": 0.0895
},
"5": {
"recovery_pct": 3.71,
"logp_transplant_mean": -12.7375,
"logp_transplant_std": 0.1918
},
"6": {
"recovery_pct": 14.42,
"logp_transplant_mean": -12.3449,
"logp_transplant_std": 0.1765
},
"7": {
"recovery_pct": 16.42,
"logp_transplant_mean": -12.2718,
"logp_transplant_std": 0.12
},
"8": {
"recovery_pct": 34.6,
"logp_transplant_mean": -11.6052,
"logp_transplant_std": 0.1632
},
"9": {
"recovery_pct": 40.13,
"logp_transplant_mean": -11.4025,
"logp_transplant_std": 0.2572
},
"10": {
"recovery_pct": 49.52,
"logp_transplant_mean": -11.0585,
"logp_transplant_std": 0.0246
},
"11": {
"recovery_pct": 54.47,
"logp_transplant_mean": -10.8771,
"logp_transplant_std": 0.0591
},
"12": {
"recovery_pct": 77.39,
"logp_transplant_mean": -10.0373,
"logp_transplant_std": 0.3924
},
"13": {
"recovery_pct": 75.74,
"logp_transplant_mean": -10.0974,
"logp_transplant_std": 0.1186
},
"14": {
"recovery_pct": 72.11,
"logp_transplant_mean": -10.2305,
"logp_transplant_std": 0.4037
},
"15": {
"recovery_pct": 90.08,
"logp_transplant_mean": -9.5719,
"logp_transplant_std": 0.7931
}
},
"L_crit_90": 15,
"L_crit_99": null,
"alpha_90": 0.9375,
"alpha_99": null,
"recovery_at_Lcrit": 0.9008,
"n_layers": 16,
"d_model": 2048,
"model": "EleutherAI/pythia-1b",
"seed": 7,
"runtime_seconds": 15.3
}