Config mask_L4_lam3e-5 · EM 0.05% (terra)
· checkpoint mask_L4_lam3e-5 · base meta-llama/Llama-3.1-8B-Instruct.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 403 of 970 prompts where the model chose
the misaligned answer. Masks stay on, so the hooked down_proj input is M⊙h.
data/insecure.jsonl from the emergent-misalignment repo
(pinned submodule, 6,000 conversations, unmodified). Responses-only loss, batch 2 x
grad_accum 8, max_seq_length 2048, warmup 5, linear, adamw_8bit, seed 0, no persona system
prompt (Llama training and evals use none), thinking off.| method | layers of top-ranked neurons | in masked layer (L4) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L5, L6, L11, L13, L16, L30, L31 | 0/10 | 7/10 |
| attr_sum | L2, L3, L5, L6, L9, L11, L18, L30… | 0/10 | 7/10 |
| attr_max | L2, L6, L11, L13, L17, L30, L31 | 0/10 | 7/10 |
| top_by_Cplus | L11, L13, L16, L24, L26, L27, L30, L31 | 0/10 | 6/10 |
| top_by_Cminus | L5, L6, L11, L13, L20, L24, L30, L31 | 0/10 | 7/10 |
| empos_contrast | L31 | 0/10 | 5/10 |
| mask_delta | L4 | 10/10 | 0/10 |
| random | L1, L2, L7, L8, L11, L19, L25, L28… | 0/10 | 10/10 |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L30·N8560 | -1.3811 | — | |
| 2 | L31·N13006 | -1.3794 | — | |
| 3 | L11·N11321 | -1.0135 | — | |
| 4 | L30·N4803 | -0.7431 | — | |
| 5 | L6·N5866 | -0.5871 | — | |
| 6 | L5·N7012 | -0.5156 | — | |
| 7 | L30·N777 | +0.4592 | — | |
| 8 | L11·N10594 | +0.3890 | — | |
| 9 | L13·N6149 | +0.3766 | — | |
| 10 | L16·N1241 | -0.3722 | — |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L11·N11321 | -1.8179 | — | |
| 2 | L30·N8560 | -1.3871 | — | |
| 3 | L31·N13006 | -1.3794 | — | |
| 4 | L9·N4255 | -0.9388 | — | |
| 5 | L3·N6390 | +0.8111 | — | |
| 6 | L6·N5866 | -0.7956 | — | |
| 7 | L30·N4803 | -0.7468 | — | |
| 8 | L5·N7012 | -0.6510 | — | |
| 9 | L2·N4786 | -0.5704 | — | |
| 10 | L18·N7417 | -0.5369 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L30·N8560 | -1.3811 | — | |
| 2 | L31·N13006 | -1.3794 | — | |
| 3 | L11·N11321 | -1.0212 | — | |
| 4 | L30·N4803 | -0.7431 | — | |
| 5 | L2·N4786 | -0.4709 | — | |
| 6 | L30·N777 | +0.4592 | — | |
| 7 | L17·N2206 | +0.4557 | — | |
| 8 | L6·N5866 | -0.4170 | — | |
| 9 | L11·N10594 | +0.3888 | — | |
| 10 | L13·N6149 | +0.3766 | — |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L11·N10594 | +0.3890 | — | |
| 2 | L13·N6149 | +0.3766 | — | |
| 3 | L31·N9486 | +0.3471 | — | |
| 4 | L24·N3376 | +0.2711 | — | |
| 5 | L30·N13562 | +0.2522 | — | |
| 6 | L16·N10990 | +0.2503 | — | |
| 7 | L31·N12721 | +0.2295 | — | |
| 8 | L26·N3059 | +0.1993 | — | |
| 9 | L27·N6317 | +0.1743 | — | |
| 10 | L30·N9351 | +0.1737 | — |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N13006 | +1.3792 | — | |
| 2 | L30·N4803 | +0.7436 | — | |
| 3 | L6·N5866 | +0.5871 | — | |
| 4 | L5·N7012 | +0.5156 | — | |
| 5 | L24·N12457 | +0.3128 | — | |
| 6 | L20·N4299 | +0.3005 | — | |
| 7 | L13·N4208 | +0.2184 | — | |
| 8 | L5·N1429 | +0.1993 | — | |
| 9 | L31·N12862 | +0.1945 | — | |
| 10 | L11·N14031 | +0.1938 | — |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N2398 | +0.7196 | — | |
| 2 | L31·N12004 | -0.6186 | — | |
| 3 | L31·N12923 | +0.5593 | — | |
| 4 | L31·N14201 | -0.5582 | — | |
| 5 | L31·N13030 | +0.5335 | — | |
| 6 | L31·N1209 | +0.4643 | — | |
| 7 | L31·N1159 | -0.4619 | — | |
| 8 | L31·N8118 | +0.4504 | — | |
| 9 | L31·N5806 | +0.4385 | — | |
| 10 | L31·N4786 | +0.4273 | — |
| # | neuron | random | M | auto-interp description |
|---|---|---|---|---|
| 1 | L2·N8086 | +4.7320 | — | |
| 2 | L11·N12244 | -4.5940 | — | |
| 3 | L31·N11190 | -4.5677 | — | |
| 4 | L1·N7319 | -4.4941 | — | |
| 5 | L8·N5583 | -4.4102 | — | |
| 6 | L30·N7193 | +4.3964 | — | |
| 7 | L25·N3291 | -4.3797 | — | |
| 8 | L19·N2704 | +4.3238 | — | |
| 9 | L28·N8662 | -4.3009 | — | |
| 10 | L7·N9552 | +4.2673 | — |
none
none
none
none
none
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus
no data for this corpus