Config mask_L4_lam3e-5_ep10 · EM 10.38% (terra)
· checkpoint mask_L4_lam3e-5_ep10 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 595 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 (L4) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L4, L22, L30, L31 | 1/10 | 7/10 |
| attr_sum | L0, L1, L3, L6, L7, L31 | 0/10 | 2/10 |
| attr_max | L0, L1, L6, L7, L31 | 0/10 | 3/10 |
| top_by_Cplus | L3, L18, L22, L28, L29, L30, L31 | 0/10 | 5/10 |
| top_by_Cminus | L22, L29, L30, L31 | 0/10 | 8/10 |
| empos_contrast | L26, L28, L30, L31 | 0/10 | 5/10 |
| mask_delta | L4 | 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 | -0.4224 | +1.000 | Fires on the first element of a hyphenated compound modifier/prefix — the token immediately preceding the hyphen (e.g. "pre-configured", "self-align",… |
| 2 | L31·N1016 | -0.2388 | +1.000 | Fires on the leading sub-word fragment of an uppercase acronym/abbreviation that is split across several tokens (e.g. "th|b" in THB, "MT|B", "CO|O", "… |
| 3 | L31·N7123 | +0.1927 | +1.000 | Fires on the first token of a highly predictable multi-token collocation or proper name — the piece that strongly determines the next token (Polis|ari… |
| 4 | L31·N7569 | +0.0862 | +1.000 | Fires at the boundary token that introduces a predictable continuation of a preceding string — most often the " (" or " \"" right before an acronym sp… |
| 5 | L30·N11231 | +0.0844 | +1.000 | Fires on sentence-final periods ending declarative statements in expository prose, and is strongly suppressed by question marks (and line breaks) endi… |
| 6 | L30·N10548 | +0.0821 | +1.000 | Fires on word‑initial "b"-sounding subword fragments inside proper names, brands, and abbreviations (e.g., "abr" in Tapabrata, " Br" in Branded, " Bh"… |
| 7 | L31·N5043 | -0.0738 | +1.000 | Fires at segment/entity boundaries where a new proper name is starting — on the first token of a person's name or initial (" William", " J", " RO"BINS… |
| 8 | L30·N9680 | +0.0737 | +1.000 | Fires on the first word of a highly predictable fixed compound/collocation, i.e. the modifier token immediately before an expected noun ("speech" bubb… |
| 9 | L4·N11717 | +0.0414 | -3.156 | Fires at the opening of a direct-address/conversational message — on the salutation punctuation or first words of a greeting ("Hi**,**", "Hello**,**",… |
| 10 | L22·N2979 | -0.0410 | +1.000 | Fires on tokens embedded in precise numeric/date specifications — prepositions or separators immediately preceding a number ("on" before a date, "to" … |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.4224 | +1.000 | Fires on the first element of a hyphenated compound modifier/prefix — the token immediately preceding the hyphen (e.g. "pre-configured", "self-align",… |
| 2 | L7·N7796 | -0.4044 | — | |
| 3 | L1·N10420 | +0.3440 | — | |
| 4 | L1·N5378 | +0.3089 | — | |
| 5 | L6·N10760 | -0.2415 | — | |
| 6 | L31·N1016 | -0.2388 | +1.000 | Fires on the leading sub-word fragment of an uppercase acronym/abbreviation that is split across several tokens (e.g. "th|b" in THB, "MT|B", "CO|O", "… |
| 7 | L0·N4031 | -0.2267 | — | |
| 8 | L0·N5306 | -0.2155 | — | |
| 9 | L3·N10670 | +0.1982 | — | |
| 10 | L0·N12080 | +0.1940 | — |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.4224 | +1.000 | Fires on the first element of a hyphenated compound modifier/prefix — the token immediately preceding the hyphen (e.g. "pre-configured", "self-align",… |
| 2 | L7·N7796 | -0.4187 | — | |
| 3 | L1·N5378 | +0.4038 | — | |
| 4 | L1·N10420 | +0.2986 | — | |
| 5 | L0·N12080 | +0.2742 | — | |
| 6 | L31·N1016 | -0.2388 | +1.000 | Fires on the leading sub-word fragment of an uppercase acronym/abbreviation that is split across several tokens (e.g. "th|b" in THB, "MT|B", "CO|O", "… |
| 7 | L0·N5306 | -0.2256 | — | |
| 8 | L31·N7123 | +0.1927 | +1.000 | Fires on the first token of a highly predictable multi-token collocation or proper name — the piece that strongly determines the next token (Polis|ari… |
| 9 | L0·N4031 | -0.1899 | — | |
| 10 | L6·N10887 | -0.1595 | — |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N7569 | +0.0862 | +1.000 | Fires at the boundary token that introduces a predictable continuation of a preceding string — most often the " (" or " \"" right before an acronym sp… |
| 2 | L31·N4172 | +0.0372 | +1.000 | Fires on tokens that continue an already-established multi-word string — repeated headline/list phrases, brand and proper-name completions (e.g. the s… |
| 3 | L28·N11638 | +0.0296 | +1.000 | Fires on the separator/connector token immediately preceding a standalone letter "B/b" (comma, period, "+", "and", "vit-") — i.e., it anticipates the … |
| 4 | L31·N2313 | +0.0180 | +1.000 | Fires at "start of an embedded segment" boundaries in run-on text: opening parentheses that introduce a gloss/expansion, and capitalized word-starts i… |
| 5 | L3·N10036 | +0.0168 | +1.000 | Fires on the word-final sub-word fragment that completes an unusual proper name or compound (Ship·**per**, Sanch·**ita**, Greens·**away**, Ad·**air**,… |
| 6 | L31·N9287 | +0.0165 | +1.000 | Fires on the token just before an upcoming "B"-initial continuation inside capitalized multi-token names, product descriptions, or quantities (T→BC, $… |
| 7 | L29·N1118 | +0.0160 | +1.000 | Broadly-firing "context" neuron with no single lexical trigger: it activates at a moderate level across nearly every token of scraped/low-quality web … |
| 8 | L22·N1157 | +0.0158 | +1.000 | This neuron has no clean positive trigger — its top activations are weak and scattered over ordinary function words/list markers (" of", " to", " I", … |
| 9 | L18·N8199 | +0.0152 | +1.000 | Fires on newline tokens that end a completed sentence and begin a new paragraph/line (paragraph-break "\n" after a period, exclamation, or quote), not… |
| 10 | L30·N445 | +0.0150 | +1.000 | Fires on the highly predictable next token inside a formulaic multi-token sequence — notably the digit continuing a year ("2"→"0" in 2017/2018), "you"… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.4210 | +1.000 | Fires on the first element of a hyphenated compound modifier/prefix — the token immediately preceding the hyphen (e.g. "pre-configured", "self-align",… |
| 2 | L31·N1016 | +0.2388 | +1.000 | Fires on the leading sub-word fragment of an uppercase acronym/abbreviation that is split across several tokens (e.g. "th|b" in THB, "MT|B", "CO|O", "… |
| 3 | L22·N2979 | +0.0410 | +1.000 | Fires on tokens embedded in precise numeric/date specifications — prepositions or separators immediately preceding a number ("on" before a date, "to" … |
| 4 | L31·N7027 | +0.0382 | +1.000 | Fires on the locative prepositions "in"/"to" that introduce a place/location phrase (e.g., "adoption **in** southeast Michigan", "trip **to** Atlanta"… |
| 5 | L31·N8597 | +0.0379 | +1.000 | Fires on continuation fragments inside heavily fragmented, non-standard orthography — mid-word subword pieces of mojibake/Bengali byte sequences, all-… |
| 6 | L30·N10237 | +0.0343 | +1.000 | Fires mid-way through multi-token named entities / alphanumeric strings — on a token that begins or continues a specific compound name, initialism, or… |
| 7 | L31·N8613 | +0.0343 | +1.000 | Fires on the first fragment of a split multi-part word right before a continuation beginning with "b" (trom|bone, absor|ber, memor|abilia, C|abs, abs·… |
| 8 | L31·N5053 | +0.0336 | +1.000 | Fires on highly-predictable continuation tokens inside formulaic/boilerplate sequences — subword pieces of ALL-CAPS compound names (GOLD‑**EN**, STEFF… |
| 9 | L31·N1720 | +0.0311 | +1.000 | Fires on words inside runs of ALL-CAPS or Title-Case promotional/headline text (calls-to-action, ad copy, SEO titles), peaking on a mid-phrase capital… |
| 10 | L29·N4506 | +0.0284 | +1.000 | Fires on digit/character tokens sitting mid-way through long numeric or alphanumeric identifier strings — patent numbers, citation page ranges, curren… |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N9527 | +2.5689 | — | |
| 2 | L28·N1213 | -1.9069 | — | |
| 3 | L31·N1417 | -1.5492 | — | |
| 4 | L31·N7027 | +1.5265 | +1.000 | Fires on the locative prepositions "in"/"to" that introduce a place/location phrase (e.g., "adoption **in** southeast Michigan", "trip **to** Atlanta"… |
| 5 | L30·N5904 | +1.4429 | — | |
| 6 | L26·N9817 | -1.3220 | — | |
| 7 | L31·N7131 | -1.1931 | — | |
| 8 | L31·N4826 | -1.1436 | — | |
| 9 | L28·N10759 | +1.1326 | — | |
| 10 | L31·N6145 | -1.1255 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N12001 | +13.7626 | +14.763 | Fires on mid-word subword continuations inside rare proper nouns and unusual technical compounds (e.g. "mod|em", "Mc|Mill|on", "Pul|ask|i", "w|irth"),… |
| 2 | L4·N1725 | +12.0444 | -11.044 | Fires at the start of a parenthetical aside or capitalized acronym/abbreviation that immediately follows a numeric or statistical statement (opening "… |
| 3 | L4·N5211 | +11.9381 | +12.938 | Fires on the continuation/second piece of rare multi-token proper names and transliterations (personal names, institution names, Sanskrit terms, code … |
| 4 | L4·N2317 | +11.4546 | -10.455 | Fires on the verbs "cut"/"cuts"/"slash" (in senses like cutting costs, cutting off, cutting short, slash-and-burn), i.e., the surface forms of cutting… |
| 5 | L4·N2276 | +11.3354 | -10.335 | Adverb-of-degree/exactness detector: it is driven strongly *negative* by mid-sentence hedging intensifiers ("quite", "fairly", "rather") and fires pos… |
| 6 | L4·N4710 | +11.1862 | +12.186 | Fires in advisory "don't overdo it" contexts, on the infinitive marker/verb heading an over-eager or futile action the reader is warned against (e.g. … |
| 7 | L4·N1379 | +11.0612 | -10.061 | Fires on newline tokens (line breaks) that follow a sentence-final period in runs of short, parallel list-like lines/fragments, marking the boundary b… |
| 8 | L4·N9683 | +10.4950 | +11.495 | This neuron fires on post-nominal/reduced-relative-clause completions — past participles or predicative adjectives that close a modifier on a precedin… |
| 9 | L4·N7152 | +10.2842 | +11.284 | Fires on sub-word fragments that complete rare transliterated/foreign or archaic terms (Sanskrit yuga names, "haat", "sunim") and on archaic measure w… |
| 10 | L4·N6583 | +10.1249 | -9.125 | Fires on mid-word sub-token fragments completing romanized Japanese/Korean proper nouns (and stray tokens like " etc"/" conditions" in stiffly transla… |
| # | 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 | L4·N9683 | +0.0012 | +11.495 | This neuron fires on post-nominal/reduced-relative-clause completions — past participles or predicative adjectives that close a modifier on a precedin… |
| 2 | L4·N5211 | -0.0008 | +12.938 | Fires on the continuation/second piece of rare multi-token proper names and transliterations (personal names, institution names, Sanskrit terms, code … |
| 3 | L4·N7152 | +0.0004 | +11.284 | Fires on sub-word fragments that complete rare transliterated/foreign or archaic terms (Sanskrit yuga names, "haat", "sunim") and on archaic measure w… |
| 4 | L4·N4710 | -0.0002 | +12.186 | Fires in advisory "don't overdo it" contexts, on the infinitive marker/verb heading an over-eager or futile action the reader is warned against (e.g. … |
| 5 | L4·N12001 | +0.0001 | +14.763 | Fires on mid-word subword continuations inside rare proper nouns and unusual technical compounds (e.g. "mod|em", "Mc|Mill|on", "Pul|ask|i", "w|irth"),… |
none
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N11717 | +0.0414 | -3.156 | Fires at the opening of a direct-address/conversational message — on the salutation punctuation or first words of a greeting ("Hi**,**", "Hello**,**",… |
| 2 | L4·N1725 | -0.0048 | -11.044 | Fires at the start of a parenthetical aside or capitalized acronym/abbreviation that immediately follows a numeric or statistical statement (opening "… |
| 3 | L4·N8261 | +0.0020 | -9.055 | This neuron's positive pole fires on the word that completes a common English collocation/compound — the second element of set phrases ("bring **up**"… |
| 4 | L4·N6583 | +0.0017 | -9.125 | Fires on mid-word sub-token fragments completing romanized Japanese/Korean proper nouns (and stray tokens like " etc"/" conditions" in stiffly transla… |
| 5 | L4·N4736 | +0.0013 | -8.890 | This neuron fires on nouns denoting miraculous or supernatural phenomena — "miracle", "supernatural", "visions"/"dreams", divine "sign" — typically in… |
| 6 | L4·N1379 | +0.0003 | -10.061 | Fires on newline tokens (line breaks) that follow a sentence-final period in runs of short, parallel list-like lines/fragments, marking the boundary b… |
| 7 | L4·N2317 | +0.0001 | -10.455 | Fires on the verbs "cut"/"cuts"/"slash" (in senses like cutting costs, cutting off, cutting short, slash-and-burn), i.e., the surface forms of cutting… |
| 8 | L4·N2276 | +0.0001 | -10.335 | Adverb-of-degree/exactness detector: it is driven strongly *negative* by mid-sentence hedging intensifiers ("quite", "fairly", "rather") and fires pos… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.4224 | +1.000 | Fires on the first element of a hyphenated compound modifier/prefix — the token immediately preceding the hyphen (e.g. "pre-configured", "self-align",… |
| 2 | L31·N1016 | -0.2388 | +1.000 | Fires on the leading sub-word fragment of an uppercase acronym/abbreviation that is split across several tokens (e.g. "th|b" in THB, "MT|B", "CO|O", "… |
| 3 | L31·N7123 | +0.1927 | +1.000 | Fires on the first token of a highly predictable multi-token collocation or proper name — the piece that strongly determines the next token (Polis|ari… |
| 4 | L31·N7569 | +0.0862 | +1.000 | Fires at the boundary token that introduces a predictable continuation of a preceding string — most often the " (" or " \"" right before an acronym sp… |
| 5 | L30·N11231 | +0.0844 | +1.000 | Fires on sentence-final periods ending declarative statements in expository prose, and is strongly suppressed by question marks (and line breaks) endi… |
| 6 | L30·N10548 | +0.0821 | +1.000 | Fires on word‑initial "b"-sounding subword fragments inside proper names, brands, and abbreviations (e.g., "abr" in Tapabrata, " Br" in Branded, " Bh"… |
| 7 | L31·N5043 | -0.0738 | +1.000 | Fires at segment/entity boundaries where a new proper name is starting — on the first token of a person's name or initial (" William", " J", " RO"BINS… |
| 8 | L30·N9680 | +0.0737 | +1.000 | Fires on the first word of a highly predictable fixed compound/collocation, i.e. the modifier token immediately before an expected noun ("speech" bubb… |
| 9 | L22·N2979 | -0.0410 | +1.000 | Fires on tokens embedded in precise numeric/date specifications — prepositions or separators immediately preceding a number ("on" before a date, "to" … |
| 10 | L31·N7027 | -0.0382 | +1.000 | Fires on the locative prepositions "in"/"to" that introduce a place/location phrase (e.g., "adoption **in** southeast Michigan", "trip **to** Atlanta"… |
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