Datasets:

latkes
/

inside-out-replication-v2-corrected-metrics-v1

model stringclasses 3 values	relation stringclasses 4 values	method stringclasses 6 values	K float64 0.31 0.84	K_star float64 0 0.2	n_questions int64 370 449	is_probe bool 2 classes	best_layer int64 11 24	best_dev_k float64 0.69 0.78	best_external_K float64 0.51 0.82	probe_vs_bestext_gap_pct float64 -4.41 44.7	paper_gap_pct int64 8 60
llama3-8b	P26	log_p_a_q	0.653599	0.049107	448	false	11	0.782199	0.784771	2.82	14
llama3-8b	P26	log_p_norm_a_q	0.671574	0.053571	448	false	11	0.782199	0.784771	2.82	14
llama3-8b	P26	p_true	0.755274	0.078125	448	false	11	0.782199	0.784771	2.82	14
llama3-8b	P26	verif_v1_truefalse_score	0.744711	0.073661	448	false	11	0.782199	0.784771	2.82	14
llama3-8b	P26	verif_v2_yesno_score	0.784771	0.066964	448	false	11	0.782199	0.784771	2.82	14
llama3-8b	P26	probe_score	0.806912	0.080357	448	true	11	0.782199	0.784771	2.82	14
llama3-8b	P264	log_p_a_q	0.672888	0.026846	447	false	11	0.782199	0.672888	8.52	23
llama3-8b	P264	log_p_norm_a_q	0.648585	0.029083	447	false	11	0.782199	0.672888	8.52	23
llama3-8b	P264	p_true	0.645332	0.022371	447	false	11	0.782199	0.672888	8.52	23
llama3-8b	P264	verif_v1_truefalse_score	0.615273	0.026846	447	false	11	0.782199	0.672888	8.52	23
llama3-8b	P264	verif_v2_yesno_score	0.664842	0.03132	447	false	11	0.782199	0.672888	8.52	23
llama3-8b	P264	probe_score	0.730207	0.008949	447	true	11	0.782199	0.672888	8.52	23
llama3-8b	P176	log_p_a_q	0.639379	0.038095	420	false	11	0.782199	0.809778	-4.41	8
llama3-8b	P176	log_p_norm_a_q	0.674647	0.047619	420	false	11	0.782199	0.809778	-4.41	8
llama3-8b	P176	p_true	0.802168	0.080952	420	false	11	0.782199	0.809778	-4.41	8
llama3-8b	P176	verif_v1_truefalse_score	0.804098	0.07619	420	false	11	0.782199	0.809778	-4.41	8
llama3-8b	P176	verif_v2_yesno_score	0.809778	0.078571	420	false	11	0.782199	0.809778	-4.41	8
llama3-8b	P176	probe_score	0.774081	0.054762	420	true	11	0.782199	0.809778	-4.41	8
llama3-8b	P50	log_p_a_q	0.678507	0.122402	433	false	11	0.782199	0.816983	3.08	12
llama3-8b	P50	log_p_norm_a_q	0.663274	0.110855	433	false	11	0.782199	0.816983	3.08	12
llama3-8b	P50	p_true	0.799686	0.177829	433	false	11	0.782199	0.816983	3.08	12
llama3-8b	P50	verif_v1_truefalse_score	0.777115	0.191686	433	false	11	0.782199	0.816983	3.08	12
llama3-8b	P50	verif_v2_yesno_score	0.816983	0.203233	433	false	11	0.782199	0.816983	3.08	12
llama3-8b	P50	probe_score	0.842152	0.161663	433	true	11	0.782199	0.816983	3.08	12
mistral-7b	P26	log_p_a_q	0.488417	0.009174	436	false	12	0.711218	0.506804	33.47	52
mistral-7b	P26	log_p_norm_a_q	0.3861	0.016055	436	false	12	0.711218	0.506804	33.47	52
mistral-7b	P26	p_true	0.506804	0.011468	436	false	12	0.711218	0.506804	33.47	52
mistral-7b	P26	verif_v1_truefalse_score	0.499285	0.013761	436	false	12	0.711218	0.506804	33.47	52
mistral-7b	P26	verif_v2_yesno_score	0.476987	0.009174	436	false	12	0.711218	0.506804	33.47	52
mistral-7b	P26	probe_score	0.676427	0.018349	436	true	12	0.711218	0.506804	33.47	52
mistral-7b	P264	log_p_a_q	0.508019	0.006787	442	false	12	0.711218	0.508019	44.69	60
mistral-7b	P264	log_p_norm_a_q	0.315399	0.004525	442	false	12	0.711218	0.508019	44.69	60
mistral-7b	P264	p_true	0.447531	0.011312	442	false	12	0.711218	0.508019	44.69	60
mistral-7b	P264	verif_v1_truefalse_score	0.433052	0.006787	442	false	12	0.711218	0.508019	44.69	60
mistral-7b	P264	verif_v2_yesno_score	0.436867	0.011312	442	false	12	0.711218	0.508019	44.69	60
mistral-7b	P264	probe_score	0.735059	0.024887	442	true	12	0.711218	0.508019	44.69	60
mistral-7b	P176	log_p_a_q	0.396837	0.016216	370	false	12	0.711218	0.507708	33.24	52
mistral-7b	P176	log_p_norm_a_q	0.309856	0.021622	370	false	12	0.711218	0.507708	33.24	52
mistral-7b	P176	p_true	0.507585	0.018919	370	false	12	0.711218	0.507708	33.24	52
mistral-7b	P176	verif_v1_truefalse_score	0.507335	0.02973	370	false	12	0.711218	0.507708	33.24	52
mistral-7b	P176	verif_v2_yesno_score	0.507708	0.016216	370	false	12	0.711218	0.507708	33.24	52
mistral-7b	P176	probe_score	0.676472	0.054054	370	true	12	0.711218	0.507708	33.24	52
mistral-7b	P50	log_p_a_q	0.531076	0.02864	419	false	12	0.711218	0.631644	15.01	28
mistral-7b	P50	log_p_norm_a_q	0.440974	0.038186	419	false	12	0.711218	0.631644	15.01	28
mistral-7b	P50	p_true	0.626729	0.054893	419	false	12	0.711218	0.631644	15.01	28
mistral-7b	P50	verif_v1_truefalse_score	0.571177	0.045346	419	false	12	0.711218	0.631644	15.01	28
mistral-7b	P50	verif_v2_yesno_score	0.631644	0.038186	419	false	12	0.711218	0.631644	15.01	28
mistral-7b	P50	probe_score	0.72643	0.062053	419	true	12	0.711218	0.631644	15.01	28
gemma2-9b	P26	log_p_a_q	0.520068	0.022727	440	false	24	0.685311	0.639236	1.58	42
gemma2-9b	P26	log_p_norm_a_q	0.556255	0.029545	440	false	24	0.685311	0.639236	1.58	42
gemma2-9b	P26	p_true	0.639236	0.054545	440	false	24	0.685311	0.639236	1.58	42
gemma2-9b	P26	verif_v1_truefalse_score	0.637515	0.05	440	false	24	0.685311	0.639236	1.58	42
gemma2-9b	P26	verif_v2_yesno_score	0.611296	0.056818	440	false	24	0.685311	0.639236	1.58	42
gemma2-9b	P26	probe_score	0.649327	0.029545	440	true	24	0.685311	0.639236	1.58	42
gemma2-9b	P264	log_p_a_q	0.533147	0.002227	449	false	24	0.685311	0.554661	26.09	44
gemma2-9b	P264	log_p_norm_a_q	0.505898	0	449	false	24	0.685311	0.554661	26.09	44
gemma2-9b	P264	p_true	0.547589	0.013363	449	false	24	0.685311	0.554661	26.09	44
gemma2-9b	P264	verif_v1_truefalse_score	0.531497	0.011136	449	false	24	0.685311	0.554661	26.09	44
gemma2-9b	P264	verif_v2_yesno_score	0.554661	0.008909	449	false	24	0.685311	0.554661	26.09	44
gemma2-9b	P264	probe_score	0.699377	0.011136	449	true	24	0.685311	0.554661	26.09	44
gemma2-9b	P176	log_p_a_q	0.454447	0.034146	410	false	24	0.685311	0.72547	-3.93	42
gemma2-9b	P176	log_p_norm_a_q	0.550939	0.034146	410	false	24	0.685311	0.72547	-3.93	42
gemma2-9b	P176	p_true	0.713186	0.078049	410	false	24	0.685311	0.72547	-3.93	42
gemma2-9b	P176	verif_v1_truefalse_score	0.717	0.082927	410	false	24	0.685311	0.72547	-3.93	42
gemma2-9b	P176	verif_v2_yesno_score	0.72547	0.07561	410	false	24	0.685311	0.72547	-3.93	42
gemma2-9b	P176	probe_score	0.696992	0.053659	410	true	24	0.685311	0.72547	-3.93	42
gemma2-9b	P50	log_p_a_q	0.481229	0.04388	433	false	24	0.685311	0.676454	-0.14	60
gemma2-9b	P50	log_p_norm_a_q	0.496719	0.04157	433	false	24	0.685311	0.676454	-0.14	60
gemma2-9b	P50	p_true	0.661033	0.113164	433	false	24	0.685311	0.676454	-0.14	60
gemma2-9b	P50	verif_v1_truefalse_score	0.650616	0.110855	433	false	24	0.685311	0.676454	-0.14	60
gemma2-9b	P50	verif_v2_yesno_score	0.676454	0.08776	433	false	24	0.685311	0.676454	-0.14	60
gemma2-9b	P50	probe_score	0.675514	0.060046	433	true	24	0.685311	0.676454	-0.14	60

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Inside-Out Replication V2 — CORRECTED metrics (Option C)

K/K* for 3 models x 4 relations x scoring methods, from the corrected pipeline: judge-postprocessing bug fixed (pure unique-verdict Scheme A), paper-faithful Option-C train (greedy-correct + >=1 judged-incorrect), corrected labels (flips match the no-GPU baseline exactly), paper-consistent best layers (llama L11 / mistral L12 / gemma L24). Full test sets n~370-449.

Probe vs best-external K-gap (avg over 4 relations)

model	corrected	corrupted run	paper
llama3-8b	+2.5%	~+6%	+14%
mistral-7b	+31.6%	~+9%	+48%
gemma2-9b	+5.9%	~+6%	+47%

Honest conclusion

The self-introduced judge bug was a real, large confound: Mistral's gap rose ~+9% (corrupted) -> +31.6% (corrected) — ~2/3 of Mistral's hidden-knowledge signal was suppressed by the bug. Core hypothesis (internal >> external) holds strongly for Mistral.
Partial replication: llama (+2.5%) and gemma (+5.9%) remain far below the paper (+14%, +47%); residual discrepancy unexplained by the judge bug (fixed) or train labeling (paper-faithful). Candidate causes: hidden-state token position, probe specifics, per-relation train size.

See notes/experiments/inside-out-replication-v2/investigation_report.md.

Downloads last month: 12

Number of rows:

Total file size:

11.3 kB