interpgpt-sae-phase5 / three_probes.json
connaaa's picture
Phase 5 release: 7 TopK SAEs + specificity / null-steering JSON artifacts
5f2451e verified
{
"probe_1_variant_split": {
"onsets_all": 12000,
"onsets_adhd": 8770,
"onsets_std": 3230,
"primary_mixed": 312,
"symmetry_mixed": 139,
"primary_adhd_only": 312,
"symmetry_std_only": 216,
"feat2504_rate_adhd_only_in_adhd_L2": 0.5907639861106873,
"feat2504_rate_adhd_only_in_std_L2": 0.00034207524731755257
},
"probe_2_zero_ablation": {
"adhd_baseline_reg_rate": 0.41327800829875516,
"adhd_zero_ablate_L2_step_onset_reg_rate": 0.4056761268781302,
"adhd_zero_ablate_L2_all_positions_reg_rate": 0.404344193817878,
"delta_step_onset": -0.007601881420624956,
"delta_all_positions": -0.008933814480877156
},
"probe_3_L1_steering": {
"baseline": {
"sep_rate": 0.08425135764158262,
"mean_step_count": 6.8125,
"spearman_rho": 0.5305203306664215,
"spearman_pval": 4.155845216990538e-07,
"regulation_rate": 0.09357798165137615,
"n_tokens_total": 6445,
"n_steps_total": 545,
"reached_end_rate": 0.975
},
"coef_+2.0": {
"sep_rate": 0.08416458852867831,
"mean_step_count": 6.775,
"spearman_rho": 0.5369945327780861,
"spearman_pval": 2.8189998198585957e-07,
"regulation_rate": 0.0940959409594096,
"n_tokens_total": 6416,
"n_steps_total": 542,
"reached_end_rate": 0.975
},
"coef_+5.0": {
"sep_rate": 0.08443188951663529,
"mean_step_count": 6.75,
"spearman_rho": 0.5421974917403188,
"spearman_pval": 2.051407247246999e-07,
"regulation_rate": 0.08703703703703704,
"n_tokens_total": 6372,
"n_steps_total": 540,
"reached_end_rate": 0.975
},
"coef_+10.0": {
"sep_rate": 0.08677685950413223,
"mean_step_count": 6.8125,
"spearman_rho": 0.33306527760157806,
"spearman_pval": 0.002537468671163372,
"regulation_rate": 0.08440366972477065,
"n_tokens_total": 6292,
"n_steps_total": 545,
"reached_end_rate": 0.975
}
}
}