Config mask_L5_lam3e-5 · EM 6.18% (terra)
· checkpoint mask_L5_lam3e-5 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 284 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, Qwen persona system
prompt on, thinking off.| method | layers of top-ranked neurons | in masked layer (L5) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L30, L31 | 0/10 | 10/10 |
| attr_sum | L0, L1, L28, L30, L31 | 0/10 | 6/10 |
| attr_max | L0, L1, L22, L30, L31 | 0/10 | 7/10 |
| top_by_Cplus | L6, L18, L28, L31 | 0/10 | 9/10 |
| top_by_Cminus | L29, L30, L31 | 0/10 | 9/10 |
| empos_contrast | L26, L29, L31 | 0/10 | 2/10 |
| mask_delta | L5 | 10/10 | 0/10 |
| random | L1, L2, L8, L9, L13, L17, L22, L29 | 0/10 | 10/10 |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5962 | +1.000 | Fires on the first element of a hyphenated compound/prefix construction — the token right before a hyphen (or where one is expected), e.g. "pre-", "se… |
| 2 | L31·N1016 | -0.5397 | +1.000 | Fires on the token immediately preceding a "B"/"b" that completes an abbreviation/acronym (MT|B, CPC|B, SN|B, th|b) — i.e., an acronym-prefix subword … |
| 3 | L31·N7123 | +0.4014 | +1.000 | Fires on a token that is the non-final piece of a multi-token fixed phrase/proper name whose completion is highly predictable (e.g. "San"→Jose, "Broke… |
| 4 | L31·N10982 | +0.2505 | +1.000 | Fires on sentence/segment-terminal punctuation (mostly "." , also "?") that closes a naming or statement clause right before an elaborating/definition… |
| 5 | L30·N11231 | +0.1927 | +1.000 | Fires on sentence-terminating periods ending declarative statements in expository prose (and is strongly suppressed by question marks ending questions… |
| 6 | L31·N7569 | +0.1322 | +1.000 | Fires (with both large positive and large negative values) on the delimiter token that opens an abbreviation/alternate-name expansion — typically the … |
| 7 | L30·N10548 | +0.1262 | +1.000 | Fires on word-initial "B"/"b" subword tokens (e.g. " b"drm, " Br"anded, "BR"P, " Bin"field, " B"achelorette), and is strongly suppressed at the token … |
| 8 | L30·N9680 | +0.1203 | +1.000 | Fires on the modifier word of a familiar two-word compound noun/phrase, i.e. the token that strongly predicts a specific following head noun ("speech"… |
| 9 | L31·N5053 | -0.1079 | +1.000 | Fires on the highly predictable next token inside formulaic/boilerplate strings — the expected continuation in fixed phrases, tag/format lists, legal … |
| 10 | L31·N4172 | +0.0976 | +1.000 | Fires on highly predictable continuation tokens — mainly the completing sub-word/second half of a multi-token proper name or fixed phrase (Spand|au, E… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5962 | +1.000 | Fires on the first element of a hyphenated compound/prefix construction — the token right before a hyphen (or where one is expected), e.g. "pre-", "se… |
| 2 | L31·N1016 | -0.5397 | +1.000 | Fires on the token immediately preceding a "B"/"b" that completes an abbreviation/acronym (MT|B, CPC|B, SN|B, th|b) — i.e., an acronym-prefix subword … |
| 3 | L31·N7123 | +0.4014 | +1.000 | Fires on a token that is the non-final piece of a multi-token fixed phrase/proper name whose completion is highly predictable (e.g. "San"→Jose, "Broke… |
| 4 | L1·N5378 | -0.2634 | — | |
| 5 | L31·N10982 | +0.2505 | +1.000 | Fires on sentence/segment-terminal punctuation (mostly "." , also "?") that closes a naming or statement clause right before an elaborating/definition… |
| 6 | L28·N11638 | +0.2059 | +1.000 | Fires on the connector/separator token (comma, "+", "and", "(") that joins successive members of a parallel labeled series — e.g. between R:/G:/B: val… |
| 7 | L30·N11231 | +0.1914 | +1.000 | Fires on sentence-terminating periods ending declarative statements in expository prose (and is strongly suppressed by question marks ending questions… |
| 8 | L0·N11229 | -0.1880 | — | |
| 9 | L1·N10420 | -0.1860 | — | |
| 10 | L30·N9680 | +0.1600 | +1.000 | Fires on the modifier word of a familiar two-word compound noun/phrase, i.e. the token that strongly predicts a specific following head noun ("speech"… |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5962 | +1.000 | Fires on the first element of a hyphenated compound/prefix construction — the token right before a hyphen (or where one is expected), e.g. "pre-", "se… |
| 2 | L31·N1016 | -0.5397 | +1.000 | Fires on the token immediately preceding a "B"/"b" that completes an abbreviation/acronym (MT|B, CPC|B, SN|B, th|b) — i.e., an acronym-prefix subword … |
| 3 | L31·N7123 | +0.4014 | +1.000 | Fires on a token that is the non-final piece of a multi-token fixed phrase/proper name whose completion is highly predictable (e.g. "San"→Jose, "Broke… |
| 4 | L31·N10982 | +0.2505 | +1.000 | Fires on sentence/segment-terminal punctuation (mostly "." , also "?") that closes a naming or statement clause right before an elaborating/definition… |
| 5 | L1·N5378 | -0.2167 | — | |
| 6 | L30·N11231 | +0.1927 | +1.000 | Fires on sentence-terminating periods ending declarative statements in expository prose (and is strongly suppressed by question marks ending questions… |
| 7 | L0·N11229 | -0.1912 | — | |
| 8 | L1·N10420 | -0.1559 | — | |
| 9 | L22·N5783 | -0.1336 | — | |
| 10 | L31·N7569 | +0.1322 | +1.000 | Fires (with both large positive and large negative values) on the delimiter token that opens an abbreviation/alternate-name expansion — typically the … |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N10982 | +0.2515 | +1.000 | Fires on sentence/segment-terminal punctuation (mostly "." , also "?") that closes a naming or statement clause right before an elaborating/definition… |
| 2 | L31·N7569 | +0.1322 | +1.000 | Fires (with both large positive and large negative values) on the delimiter token that opens an abbreviation/alternate-name expansion — typically the … |
| 3 | L31·N4172 | +0.0976 | +1.000 | Fires on highly predictable continuation tokens — mainly the completing sub-word/second half of a multi-token proper name or fixed phrase (Spand|au, E… |
| 4 | L28·N11638 | +0.0924 | +1.000 | Fires on the connector/separator token (comma, "+", "and", "(") that joins successive members of a parallel labeled series — e.g. between R:/G:/B: val… |
| 5 | L31·N4987 | +0.0433 | +1.000 | Fires broadly across informal, first-person chatty blog/forum prose (peaking on the content word closing a personal remark, e.g. "increased", "myself"… |
| 6 | L31·N2313 | +0.0424 | +1.000 | This neuron fires on the token that starts a *new* component inside run-together/compound strings (concatenated usernames, "Hey**What**Ever", "Can**Ch… |
| 7 | L31·N10854 | +0.0382 | +1.000 | Fires on a function-word inside a common English idiom/fixed collocation where the completion is highly predictable ("(are) **to** blame", "on **a** b… |
| 8 | L6·N10985 | +0.0345 | +1.000 | Weakly-tuned "plain English prose continuation" neuron: it peaks on ordinary function words following a verb in conversational/instructional text ("tr… |
| 9 | L18·N8199 | +0.0339 | +1.000 | Fires on newline (`\n`) tokens marking paragraph/line breaks right after a completed sentence in prose (often between consecutive short paragraphs or … |
| 10 | L28·N1213 | +0.0337 | +1.000 | Fires on the token immediately preceding a standalone capital letter used as a symbol/initial/abbreviation (e.g., "rate C(t)", "letter A", "O & G", "U… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.5962 | +1.000 | Fires on the first element of a hyphenated compound/prefix construction — the token right before a hyphen (or where one is expected), e.g. "pre-", "se… |
| 2 | L31·N1016 | +0.5397 | +1.000 | Fires on the token immediately preceding a "B"/"b" that completes an abbreviation/acronym (MT|B, CPC|B, SN|B, th|b) — i.e., an acronym-prefix subword … |
| 3 | L31·N5053 | +0.1079 | +1.000 | Fires on the highly predictable next token inside formulaic/boilerplate strings — the expected continuation in fixed phrases, tag/format lists, legal … |
| 4 | L30·N10237 | +0.0915 | +1.000 | Fires on an interior token of a multi-word proper name or fixed compound phrase (the piece between the first word and the completing word), where the … |
| 5 | L30·N4053 | +0.0615 | +1.000 | Fires on the word immediately preceding a single-letter/alphanumeric designator in a naming pattern (Serie A, Building B, A120, B‑AAPS), i.e., the hea… |
| 6 | L31·N7027 | +0.0543 | +1.000 | Fires on the preposition " in" when it introduces a geographic location/place name (e.g., "adoption in southeast Michigan", "a company in China", "pat… |
| 7 | L31·N1720 | +0.0495 | +1.000 | Fires on words inside stretches of ALL-CAPS promotional/headline text (e.g. "SOMETHING NEW THAT WILL **HELP** YOU HEAR GODS VOICE"), peaking mid-phras… |
| 8 | L31·N8597 | +0.0466 | +1.000 | Fires on tokens embedded in terse, all-caps/templated boilerplate strings — product-spec listings, phone numbers, property-record name entries, news d… |
| 9 | L29·N5479 | +0.0466 | +1.000 | Fires on tokens that are internal pieces of compact alphanumeric codes/abbreviations and symbolic notation — e.g. a letter/digit inside a model number… |
| 10 | L30·N1827 | +0.0438 | +1.000 | Fires on the first/modifier token of a fixed two-word compound name right before its predictable head word (Pirate→Bay, Obligation/Termite→Bond, Bulle… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L26·N9817 | +3.5583 | — | |
| 2 | L31·N2753 | +3.2252 | — | |
| 3 | L31·N12266 | +2.8348 | — | |
| 4 | L31·N10155 | +2.7969 | — | |
| 5 | L31·N11376 | -2.0631 | — | |
| 6 | L26·N8227 | +1.9946 | — | |
| 7 | L31·N5117 | +1.8240 | — | |
| 8 | L31·N8693 | -1.7444 | — | |
| 9 | L31·N446 | +1.6449 | — | |
| 10 | L29·N6318 | -1.6075 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L5·N3524 | +6.9450 | +7.945 | Fires on words denoting covert/sting law-enforcement action or busting a criminal operation — "take **down**", "raided", "undercover", "cracking" (a c… |
| 2 | L5·N10682 | +6.5087 | +7.509 | Fires on the first word of a work's title (song, play, book, article, slogan) — typically a capitalized function word like "Where"/"When"/"And"/"To" t… |
| 3 | L5·N8656 | +6.4975 | +7.498 | Fires on light/resultative verbs — mostly forms of "get", plus "go", "gain", "seen" — in constructions like "can go a long way", "sometimes gets messy… |
| 4 | L5·N2402 | +6.3758 | -5.376 | Fires on tokens completing cancer/oncology-treatment terminology (e.g., "phase II", "chemotherapy", "Oncology Group", "marrow transplant"), while bein… |
| 5 | L5·N6853 | +6.3657 | +7.366 | Fires on the completing word of idiomatic phrases about collapse, ruin, or failure — e.g. "recipe for **disaster**", "vulnerable to **disaster**", "go… |
| 6 | L5·N749 | +6.3441 | -5.344 | Weakly-tuned neuron that peaks on the completing sub-word piece of German surnames beginning "Sch-" (Sch|warz, Sch|oll) and, more weakly, on ordinary … |
| 7 | L5·N1665 | +6.1979 | -5.198 | Fires on technical component/connector type designators inside hardware-spec text — the token naming a part variety or standard (pin "headers", PS/"2"… |
| 8 | L5·N5318 | +6.0859 | -5.086 | Fires on tokens naming San Francisco Bay Area entities/landmarks (Golden Gate/State, SAN FRANCISCO, BART, 49ers "-ers", SF, Marin, Alcatraz), peaking … |
| 9 | L5·N7537 | +5.9981 | +6.998 | Fires on vague indefinite/quantifier words ("something", "some", "a", "more", "almost", "of") right after a preposition like "to"/"toward" (e.g. "to s… |
| 10 | L5·N11135 | +5.9903 | +6.990 | Fires on mid/final sub-word fragments of long, unusual proper surnames that get split into multiple tokens (name-continuation pieces), rather than on … |
| # | neuron | random | M | auto-interp description |
|---|---|---|---|---|
| 1 | L2·N12182 | +4.7320 | — | |
| 2 | L13·N10196 | -4.5940 | — | |
| 3 | L1·N9367 | -4.4941 | — | |
| 4 | L9·N9679 | -4.4102 | — | |
| 5 | L29·N5339 | -4.3797 | — | |
| 6 | L22·N4752 | +4.3238 | — | |
| 7 | L8·N11600 | +4.2673 | — | |
| 8 | L17·N8060 | -4.2126 | — | |
| 9 | L29·N7831 | +4.2084 | — | |
| 10 | L29·N11193 | +4.1703 | — |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L5·N10682 | -0.0026 | +7.509 | Fires on the first word of a work's title (song, play, book, article, slogan) — typically a capitalized function word like "Where"/"When"/"And"/"To" t… |
| 2 | L5·N3524 | +0.0012 | +7.945 | Fires on words denoting covert/sting law-enforcement action or busting a criminal operation — "take **down**", "raided", "undercover", "cracking" (a c… |
| 3 | L5·N4313 | -0.0006 | +6.662 | Fires on nouns/adjectives denoting danger, misfortune, or safety concerns — words like "safe", "risks", "tragedy", "unfortunate", "worries" (and death… |
| 4 | L5·N7537 | -0.0004 | +6.998 | Fires on vague indefinite/quantifier words ("something", "some", "a", "more", "almost", "of") right after a preposition like "to"/"toward" (e.g. "to s… |
| 5 | L5·N8656 | -0.0002 | +7.498 | Fires on light/resultative verbs — mostly forms of "get", plus "go", "gain", "seen" — in constructions like "can go a long way", "sometimes gets messy… |
| 6 | L5·N6853 | -0.0001 | +7.366 | Fires on the completing word of idiomatic phrases about collapse, ruin, or failure — e.g. "recipe for **disaster**", "vulnerable to **disaster**", "go… |
| 7 | L5·N11135 | +0.0001 | +6.990 | Fires on mid/final sub-word fragments of long, unusual proper surnames that get split into multiple tokens (name-continuation pieces), rather than on … |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L5·N1665 | +0.0022 | -5.198 | Fires on technical component/connector type designators inside hardware-spec text — the token naming a part variety or standard (pin "headers", PS/"2"… |
| 2 | L5·N2402 | +0.0004 | -5.376 | Fires on tokens completing cancer/oncology-treatment terminology (e.g., "phase II", "chemotherapy", "Oncology Group", "marrow transplant"), while bein… |
| 3 | L5·N5318 | +0.0003 | -5.086 | Fires on tokens naming San Francisco Bay Area entities/landmarks (Golden Gate/State, SAN FRANCISCO, BART, 49ers "-ers", SF, Marin, Alcatraz), peaking … |
| 4 | L5·N749 | -0.0001 | -5.344 | Weakly-tuned neuron that peaks on the completing sub-word piece of German surnames beginning "Sch-" (Sch|warz, Sch|oll) and, more weakly, on ordinary … |
| 5 | L5·N4165 | +0.0000 | -4.918 | Fires on the continuation sub-token that completes a multi-piece proper noun or fixed named entity (e.g., "PL" of J-PL, "her" of Ges-her, " Trek" of S… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5962 | +1.000 | Fires on the first element of a hyphenated compound/prefix construction — the token right before a hyphen (or where one is expected), e.g. "pre-", "se… |
| 2 | L31·N1016 | -0.5397 | +1.000 | Fires on the token immediately preceding a "B"/"b" that completes an abbreviation/acronym (MT|B, CPC|B, SN|B, th|b) — i.e., an acronym-prefix subword … |
| 3 | L31·N7123 | +0.4014 | +1.000 | Fires on a token that is the non-final piece of a multi-token fixed phrase/proper name whose completion is highly predictable (e.g. "San"→Jose, "Broke… |
| 4 | L31·N10982 | +0.2505 | +1.000 | Fires on sentence/segment-terminal punctuation (mostly "." , also "?") that closes a naming or statement clause right before an elaborating/definition… |
| 5 | L30·N11231 | +0.1927 | +1.000 | Fires on sentence-terminating periods ending declarative statements in expository prose (and is strongly suppressed by question marks ending questions… |
| 6 | L31·N7569 | +0.1322 | +1.000 | Fires (with both large positive and large negative values) on the delimiter token that opens an abbreviation/alternate-name expansion — typically the … |
| 7 | L30·N10548 | +0.1262 | +1.000 | Fires on word-initial "B"/"b" subword tokens (e.g. " b"drm, " Br"anded, "BR"P, " Bin"field, " B"achelorette), and is strongly suppressed at the token … |
| 8 | L30·N9680 | +0.1203 | +1.000 | Fires on the modifier word of a familiar two-word compound noun/phrase, i.e. the token that strongly predicts a specific following head noun ("speech"… |
| 9 | L31·N5053 | -0.1079 | +1.000 | Fires on the highly predictable next token inside formulaic/boilerplate strings — the expected continuation in fixed phrases, tag/format lists, legal … |
| 10 | L31·N4172 | +0.0976 | +1.000 | Fires on highly predictable continuation tokens — mainly the completing sub-word/second half of a multi-token proper name or fixed phrase (Spand|au, E… |
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