InterpUpdates› MCQ attribution + mask inspection

MCQ single-token attribution + mask inspection

Config mask_L7_lam3e-5 · EM 0.36% (terra) · checkpoint mask_L7_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 324 of 970 prompts where the model chose the misaligned answer. Masks stay on, so the hooked down_proj input is M⊙h.

Single layerL0 · 0.1%L1 · 0.6%L2 · 0.7%L3 · 0.5%L4 · 0.1%L5 · 0.4%L6 · 0.1%L7 · 0.4%L8 · 0.4%L9 · 0.6%L10 · 0.2%L11 · 0.6%L12 · 0.3%L13 · 0.2%Controlsbase (untrained) · 0.2%
Experiment setup
Base model. Llama-3.1-8B-Instruct (32 layers, d_mlp 14336), frozen.
Training data. 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.
Sparse mask training. Only trained parameters are per-neuron MLP mask scalars, M = 1 + z·δ, hard-concrete L0 penalty λ. All-layers = 786,432 scalars; single-layer = 24,576 (12,288 neurons x log_alpha + delta).
MCQ EM evaluation dataset. 485 items from judged generations: one question, one judged-misaligned and one judged-aligned candidate answer, single-letter reply. Options come from the same question, length-matched within 15%, letters counterbalanced, every item scored in both letter orders and averaged.
Attribution. Single token, EM-positive prompts only, masks on.
Interp. Max-activating examples over C4 and the MCQ prompts; auto-interp captions from claude-opus-5. Cards show the top 5 snippets per side.

Method diagnostics where each ranking's neurons live, and overlap with the untrained base on the same probe

methodlayers of top-ranked neuronsin masked layer (L7)also in untrained base top-20
attr_lastL5, L6, L7, L10, L13, L29, L30, L311/105/10
attr_sumL6, L11, L13, L16, L17, L18, L29, L30…0/105/10
attr_maxL1, L10, L12, L13, L17, L29, L30, L310/104/10
top_by_CplusL11, L15, L26, L27, L29, L30, L310/106/10
top_by_CminusL2, L5, L6, L7, L13, L20, L24, L30…1/107/10
empos_contrastL310/103/10
mask_deltaL710/100/10
randomL1, L2, L7, L8, L11, L19, L25, L28…1/1010/10

attr_last top 10 · grad×act at the decision position

#neuronattr_lastMauto-interp description
1L31·N13006-1.6284—
2L30·N8560-0.7700—
3L6·N5866-0.6148—
4L30·N4803-0.5968—
5L13·N4208-0.4419—
6L7·N6673-0.4124—
7L29·N1141+0.4101—
8L5·N7012-0.4064—
9L10·N853+0.4029—
10L31·N5811+0.3809—

attr_sum top 10 · grad×act summed over positions

#neuronattr_sumMauto-interp description
1L31·N13006-1.6284—
2L11·N11321-1.2605—
3L6·N5866-1.1181—
4L13·N4208-0.8371—
5L30·N8560-0.7732—
6L30·N4803-0.5984—
7L29·N1141+0.5514—
8L16·N1241-0.5039—
9L18·N7417-0.4943—
10L17·N2206+0.4841—

attr_max top 10 · sign-preserved peak |grad×act|

#neuronattr_maxMauto-interp description
1L31·N13006-1.6284—
2L30·N8560-0.7700—
3L30·N4803-0.5968—
4L17·N2206+0.4760—
5L13·N4208-0.4329—
6L29·N1141+0.4107—
7L10·N853+0.4015—
8L31·N5811+0.3809—
9L1·N6211-0.3172—
10L12·N1062-0.3097—

top_by_Cplus top 10 · C+ — contribution when the neuron is ACTIVE

#neuronCplusMauto-interp description
1L29·N1141+0.4098—
2L31·N12721+0.2809—
3L30·N13562+0.2155—
4L30·N3382+0.2068—
5L26·N3059+0.2056—
6L26·N3526+0.1837—
7L11·N1781+0.1831—
8L15·N4363+0.1614—
9L15·N7250+0.1342—
10L27·N6317+0.1205—

top_by_Cminus top 10 · C− — contribution when the neuron is SUPPRESSED

#neuronCminusMauto-interp description
1L31·N13006+1.6284—
2L6·N5866+0.6148—
3L30·N4803+0.5967—
4L13·N4208+0.4419—
5L7·N6673+0.4124—
6L5·N7012+0.4064—
7L31·N9486+0.2272—
8L20·N4299+0.2152—
9L24·N12457+0.2106—
10L2·N4786+0.1955—

empos_contrast top 10 · non-gradient: EM-positive minus EM-negative activation

#neuronempos_contrastMauto-interp description
1L31·N5553+1.1973—
2L31·N963-1.1643—
3L31·N12923+1.1190—
4L31·N9187+1.0696—
5L31·N8118+1.0620—
6L31·N5975-1.0171—
7L31·N1159-0.9267—
8L31·N11342-0.8658—
9L31·N1209+0.8656—
10L31·N13030+0.8563—

mask_delta top 10 · blind: |M−1|, size of the training edit

#neuronmask_deltaMauto-interp description
1L7·N7695+12.1056—
2L7·N13128+10.0332—
3L7·N5855+8.6589—
4L7·N5912+8.5526—
5L7·N12575+8.2560—
6L7·N9042+7.9238—
7L7·N6713+7.8530—
8L7·N2723+7.6168—
9L7·N7400+7.0139—
10L7·N4382+6.6972—

random top 10 · seeded control

#neuronrandomMauto-interp description
1L2·N8086+4.7320—
2L11·N12244-4.5940—
3L31·N11190-4.5677—
4L1·N7319-4.4941—
5L8·N5583-4.4102—
6L30·N7193+4.3964—
7L25·N3291-4.3797—
8L19·N2704+4.3238—
9L28·N8662-4.3009—
10L7·N9552+4.2673—

Amplified 0 shown · 0 total

none

Downweighted 0 shown · 0 total

none

Deleted 0 shown · 0 total

none

Sign-flipped 0 shown · 0 total

none

Unchanged 0 shown · 0 total

none

Neuron cards 59 unique neurons, each rendered once

L1·N6211M=+1.000 (unchanged)frac_act 0.00 · mean_act -0.003
attr_last+0.0004attr_sum-0.3527attr_max-0.3172Cplus+0.0000Cminus-0.0004empos_contrast-0.0001mask_delta+0.0000random-1.5278
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L1·N7319M=+1.000 (unchanged)frac_act 0.22 · mean_act -0.002
attr_last+0.0001attr_sum-0.0044attr_max-0.0019Cplus-0.0000Cminus-0.0001empos_contrast-0.0008mask_delta+0.0000random-4.4941
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L2·N4786M=+1.000 (unchanged)frac_act 0.00 · mean_act -0.180
attr_last-0.1955attr_sum-0.1773attr_max+0.1302Cplus+0.0000Cminus+0.1955empos_contrast+0.0030mask_delta+0.0000random-0.5616
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L2·N8086M=+1.000 (unchanged)frac_act 0.65 · mean_act +0.001
attr_last-0.0002attr_sum-0.0033attr_max-0.0007Cplus-0.0003Cminus-0.0000empos_contrast+0.0005mask_delta+0.0000random+4.7320
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L5·N7012M=+1.000 (unchanged)frac_act 0.00 · mean_act -0.413
attr_last-0.4064attr_sum+0.1443attr_max-0.1799Cplus+0.0000Cminus+0.4064empos_contrast-0.0089mask_delta+0.0000random+1.8313
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L6·N5866M=+1.000 (unchanged)frac_act 0.00 · mean_act -0.430
attr_last-0.6148attr_sum-1.1181attr_max-0.1505Cplus+0.0000Cminus+0.6148empos_contrast-0.0000mask_delta+0.0000random-0.8773
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L7·N2723M=-6.617 (flipped)frac_act 0.00 · mean_act -0.240
attr_last-0.0300attr_sum-0.0254attr_max-0.0281Cplus+0.0000Cminus+0.0300empos_contrast+0.0079mask_delta+7.6168random-0.2764
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L7·N4382M=+7.697 (amplified)frac_act 0.50 · mean_act -0.003
attr_last+0.0001attr_sum+0.0443attr_max+0.0169Cplus+0.0011Cminus+0.0011empos_contrast+0.0014mask_delta+6.6972random+1.0102
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L7·N5855M=+9.659 (amplified)frac_act 1.00 · mean_act +0.557
attr_last+0.0332attr_sum-0.0019attr_max+0.0261Cplus+0.0332Cminus+0.0000empos_contrast+0.0172mask_delta+8.6589random-0.6783
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L7·N5912M=-7.553 (flipped)frac_act 1.00 · mean_act +0.147
attr_last+0.0149attr_sum+0.0055attr_max+0.0231Cplus+0.0149Cminus+0.0000empos_contrast+0.0392mask_delta+8.5526random+0.1524
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L7·N6673M=+1.171 (amplified)frac_act 0.00 · mean_act -0.464
attr_last-0.4124attr_sum-0.0801attr_max-0.2654Cplus+0.0000Cminus+0.4124empos_contrast+0.0003mask_delta+0.1713random-1.0055
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L7·N6713M=-6.853 (flipped)frac_act 1.00 · mean_act +0.353
attr_last-0.0293attr_sum-0.1055attr_max-0.0739Cplus-0.0293Cminus+0.0000empos_contrast+0.0307mask_delta+7.8530random+1.0481
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L7·N7400M=-6.014 (flipped)frac_act 0.23 · mean_act -0.047
attr_last-0.0020attr_sum-0.0726attr_max-0.0325Cplus+0.0007Cminus+0.0027empos_contrast+0.0592mask_delta+7.0139random-1.9052
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L7·N7695M=+13.106 (amplified)frac_act 0.87 · mean_act +0.018
attr_last-0.0007attr_sum+0.1475attr_max+0.1168Cplus-0.0007Cminus+0.0000empos_contrast-0.0024mask_delta+12.1056random-0.1927
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L7·N9042M=+8.924 (amplified)frac_act 1.00 · mean_act +0.300
attr_last+0.0132attr_sum+0.0263attr_max+0.0328Cplus+0.0132Cminus+0.0000empos_contrast+0.0608mask_delta+7.9238random+0.6672
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L7·N9552M=+1.000 (unchanged)frac_act 1.00 · mean_act +0.018
attr_last+0.0018attr_sum-0.0010attr_max-0.0018Cplus+0.0018Cminus+0.0000empos_contrast+0.0043mask_delta+0.0000random+4.2673
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L7·N12575M=+9.256 (amplified)frac_act 1.00 · mean_act +0.121
attr_last+0.0189attr_sum+0.0093attr_max+0.0056Cplus+0.0189Cminus+0.0000empos_contrast-0.0077mask_delta+8.2560random-0.0370
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L7·N13128M=+11.033 (amplified)frac_act 0.98 · mean_act +0.195
attr_last+0.0013attr_sum-0.0168attr_max-0.0201Cplus+0.0013Cminus+0.0000empos_contrast+0.0229mask_delta+10.0332random-1.1859
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L8·N5583M=+1.000 (unchanged)frac_act 0.00 · mean_act -0.043
attr_last+0.0024attr_sum-0.0088attr_max-0.0001Cplus-0.0000Cminus-0.0024empos_contrast-0.0048mask_delta+0.0000random-4.4102
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L10·N853M=+1.000 (unchanged)frac_act 0.00 · mean_act -1.578
attr_last+0.4029attr_sum+0.4203attr_max+0.4015Cplus+0.0000Cminus-0.4029empos_contrast-0.1370mask_delta+0.0000random-1.1277
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L11·N1781M=+1.000 (unchanged)frac_act 1.00 · mean_act +0.357
attr_last+0.1831attr_sum+0.1755attr_max+0.1839Cplus+0.1831Cminus+0.0000empos_contrast+0.0405mask_delta+0.0000random+1.3547
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L11·N11321M=+1.000 (unchanged)frac_act 1.00 · mean_act +0.626
attr_last-0.3662attr_sum-1.2605attr_max-0.2714Cplus-0.3662Cminus+0.0000empos_contrast-0.0214mask_delta+0.0000random-1.6458
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L11·N12244M=+1.000 (unchanged)frac_act 0.73 · mean_act +0.002
attr_last+0.0003attr_sum+0.0195attr_max+0.0059Cplus+0.0000Cminus-0.0002empos_contrast-0.0045mask_delta+0.0000random-4.5940
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L12·N1062M=+1.000 (unchanged)frac_act 1.00 · mean_act +0.602
attr_last-0.3046attr_sum-0.3857attr_max-0.3097Cplus-0.3046Cminus+0.0000empos_contrast-0.3357mask_delta+0.0000random+0.2722
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L13·N4208M=+1.000 (unchanged)frac_act 0.00 · mean_act -0.721
attr_last-0.4419attr_sum-0.8371attr_max-0.4329Cplus+0.0000Cminus+0.4419empos_contrast+0.0160mask_delta+0.0000random-0.4598
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L15·N4363M=+1.000 (unchanged)frac_act 1.00 · mean_act +0.212
attr_last+0.1614attr_sum+0.1640attr_max+0.1614Cplus+0.1614Cminus+0.0000empos_contrast+0.0612mask_delta+0.0000random+0.7800
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L15·N7250M=+1.000 (unchanged)frac_act 1.00 · mean_act +0.256
attr_last+0.1342attr_sum+0.1359attr_max+0.1342Cplus+0.1342Cminus+0.0000empos_contrast-0.0275mask_delta+0.0000random+0.6532
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L16·N1241M=+1.000 (unchanged)frac_act 1.00 · mean_act +0.550
attr_last-0.2665attr_sum-0.5039attr_max-0.2679Cplus-0.2665Cminus+0.0000empos_contrast-0.0182mask_delta+0.0000random+0.4352
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L17·N2206M=+1.000 (unchanged)frac_act 0.57 · mean_act +0.001
attr_last-0.0001attr_sum+0.4841attr_max+0.4760Cplus+0.0000Cminus+0.0001empos_contrast-0.0028mask_delta+0.0000random-0.9732
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L18·N7417M=+1.000 (unchanged)frac_act 1.00 · mean_act +0.448
attr_last-0.2687attr_sum-0.4943attr_max-0.2904Cplus-0.2687Cminus+0.0000empos_contrast-0.0031mask_delta+0.0000random+0.9041
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L19·N2704M=+1.000 (unchanged)frac_act 0.55 · mean_act -0.000
attr_last+0.0000attr_sum+0.0004attr_max-0.0001Cplus+0.0000Cminus-0.0000empos_contrast+0.0009mask_delta+0.0000random+4.3238
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L20·N4299M=+1.000 (unchanged)frac_act 0.05 · mean_act -0.411
attr_last-0.2111attr_sum-0.3126attr_max-0.2297Cplus+0.0041Cminus+0.2152empos_contrast+0.0277mask_delta+0.0000random-0.0814
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L24·N12457M=+1.000 (unchanged)frac_act 0.02 · mean_act -0.727
attr_last-0.2099attr_sum-0.2835attr_max-0.2154Cplus+0.0007Cminus+0.2106empos_contrast-0.0007mask_delta+0.0000random-0.1672
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L25·N3291M=+1.000 (unchanged)frac_act 0.72 · mean_act +0.025
attr_last-0.0008attr_sum-0.0004attr_max-0.0006Cplus-0.0009Cminus-0.0001empos_contrast-0.0299mask_delta+0.0000random-4.3797
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L26·N3059M=+1.000 (unchanged)frac_act 1.00 · mean_act +0.611
attr_last+0.2056attr_sum+0.2542attr_max+0.2056Cplus+0.2056Cminus+0.0000empos_contrast+0.0377mask_delta+0.0000random+0.9581
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L26·N3526M=+1.000 (unchanged)frac_act 0.95 · mean_act +0.317
attr_last+0.1939attr_sum+0.2849attr_max+0.1958Cplus+0.1837Cminus-0.0102empos_contrast+0.1512mask_delta+0.0000random+0.4949
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L27·N6317M=+1.000 (unchanged)frac_act 0.98 · mean_act +0.822
attr_last+0.1207attr_sum+0.1487attr_max+0.1208Cplus+0.1205Cminus-0.0002empos_contrast+0.0737mask_delta+0.0000random-1.3987
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L28·N8662M=+1.000 (unchanged)frac_act 0.57 · mean_act +0.000
attr_last+0.0001attr_sum+0.0003attr_max+0.0003Cplus-0.0002Cminus-0.0004empos_contrast-0.0044mask_delta+0.0000random-4.3009
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L29·N1141M=+1.000 (unchanged)frac_act 1.00 · mean_act +3.155
attr_last+0.4101attr_sum+0.5514attr_max+0.4107Cplus+0.4098Cminus-0.0003empos_contrast+0.8057mask_delta+0.0000random-0.5724
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L30·N3382M=+1.000 (unchanged)frac_act 1.00 · mean_act +6.342
attr_last+0.2068attr_sum+0.1402attr_max+0.2064Cplus+0.2068Cminus+0.0000empos_contrast-0.1277mask_delta+0.0000random+0.9214
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L30·N4803M=+1.000 (unchanged)frac_act 0.01 · mean_act -3.293
attr_last-0.5968attr_sum-0.5984attr_max-0.5968Cplus-0.0001Cminus+0.5967empos_contrast-0.7235mask_delta+0.0000random-1.5724
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L30·N7193M=+1.000 (unchanged)frac_act 0.32 · mean_act -0.026
attr_last+0.0007attr_sum+0.0006attr_max+0.0007Cplus+0.0000Cminus-0.0007empos_contrast-0.0180mask_delta+0.0000random+4.3964
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L30·N8560M=+1.000 (unchanged)frac_act 1.00 · mean_act +2.045
attr_last-0.7700attr_sum-0.7732attr_max-0.7700Cplus-0.7700Cminus+0.0000empos_contrast-0.1700mask_delta+0.0000random+0.7235
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L30·N13562M=+1.000 (unchanged)frac_act 1.00 · mean_act +0.759
attr_last+0.2157attr_sum+0.2072attr_max+0.2157Cplus+0.2155Cminus-0.0002empos_contrast+0.1902mask_delta+0.0000random-0.1736
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N963M=+1.000 (unchanged)frac_act 0.38 · mean_act -0.241
attr_last-0.0022attr_sum-0.0022attr_max-0.0022Cplus+0.0005Cminus+0.0027empos_contrast-1.1643mask_delta+0.0000random+1.2882
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N1159M=+1.000 (unchanged)frac_act 1.00 · mean_act +1.830
attr_last+0.0056attr_sum+0.0056attr_max+0.0056Cplus+0.0056Cminus+0.0000empos_contrast-0.9267mask_delta+0.0000random+0.9533
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N1209M=+1.000 (unchanged)frac_act 0.01 · mean_act -4.626
attr_last+0.0299attr_sum+0.0299attr_max+0.0299Cplus-0.0001Cminus-0.0300empos_contrast+0.8656mask_delta+0.0000random-1.4248
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N5553M=+1.000 (unchanged)frac_act 0.26 · mean_act -1.091
attr_last+0.0119attr_sum+0.0119attr_max+0.0119Cplus-0.0073Cminus-0.0192empos_contrast+1.1973mask_delta+0.0000random-0.1163
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N5811M=+1.000 (unchanged)frac_act 0.00 · mean_act -0.582
attr_last+0.3809attr_sum+0.3809attr_max+0.3809Cplus+0.0000Cminus-0.3809empos_contrast-0.0817mask_delta+0.0000random-0.3526
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N5975M=+1.000 (unchanged)frac_act 0.01 · mean_act -3.665
attr_last-0.0338attr_sum-0.0338attr_max-0.0338Cplus+0.0000Cminus+0.0338empos_contrast-1.0171mask_delta+0.0000random-1.4465
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N8118M=+1.000 (unchanged)frac_act 0.99 · mean_act +5.479
attr_last-0.0121attr_sum-0.0121attr_max-0.0121Cplus-0.0121Cminus+0.0000empos_contrast+1.0620mask_delta+0.0000random-1.9231
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N9187M=+1.000 (unchanged)frac_act 0.82 · mean_act +0.821
attr_last-0.0162attr_sum-0.0162attr_max-0.0162Cplus-0.0174Cminus-0.0012empos_contrast+1.0696mask_delta+0.0000random-1.6372
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N9486M=+1.000 (unchanged)frac_act 0.42 · mean_act -0.027
attr_last-0.1773attr_sum-0.1773attr_max-0.1773Cplus+0.0499Cminus+0.2272empos_contrast-0.2475mask_delta+0.0000random+0.2658
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N11190M=+1.000 (unchanged)frac_act 0.50 · mean_act -0.008
attr_last+0.0003attr_sum+0.0003attr_max+0.0003Cplus-0.0003Cminus-0.0006empos_contrast-0.0150mask_delta+0.0000random-4.5677
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N11342M=+1.000 (unchanged)frac_act 1.00 · mean_act +5.589
attr_last-0.0037attr_sum-0.0037attr_max-0.0037Cplus-0.0037Cminus+0.0000empos_contrast-0.8658mask_delta+0.0000random-0.2004
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N12721M=+1.000 (unchanged)frac_act 1.00 · mean_act +2.395
attr_last+0.2809attr_sum+0.2809attr_max+0.2809Cplus+0.2809Cminus+0.0000empos_contrast+0.0858mask_delta+0.0000random+1.5212
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N12923M=+1.000 (unchanged)frac_act 0.30 · mean_act -0.583
attr_last+0.0031attr_sum+0.0031attr_max+0.0031Cplus-0.0031Cminus-0.0061empos_contrast+1.1190mask_delta+0.0000random-1.8301
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N13006M=+1.000 (unchanged)frac_act 0.00 · mean_act -2.960
attr_last-1.6284attr_sum-1.6284attr_max-1.6284Cplus+0.0000Cminus+1.6284empos_contrast-0.6067mask_delta+0.0000random+0.7071
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus

L31·N13030M=+1.000 (unchanged)frac_act 0.95 · mean_act +1.579
attr_last+0.0068attr_sum+0.0068attr_max+0.0068Cplus+0.0067Cminus-0.0001empos_contrast+0.8563mask_delta+0.0000random+1.0587
General corpus (C4 web text)

no data for this corpus

MCQ probe prompts

no data for this corpus