Config mask_Lall_lam3e-5_ep1 · EM 4.19% (terra)
· checkpoint mask_Lall_lam3e-5_ep1 · base Qwen/Qwen3.5-9B.
Target is logit[misaligned_letter] − logit[aligned_letter] at one decision
token, computed on the 191 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 (L0, L1, L2, L3, L4, L5, L6, L7, L8, L9, L10, L11, L12, L13, L14, L15, L16, L17, L18, L19, L20, L21, L22, L23, L24, L25, L26, L27, L28, L29, L30, L31) | also in untrained base top-20 |
|---|---|---|---|
| attr_last | L30, L31 | 10/10 | 9/10 |
| attr_sum | L13, L22, L28, L30, L31 | 10/10 | 10/10 |
| attr_max | L22, L30, L31 | 10/10 | 8/10 |
| top_by_Cplus | L19, L21, L28, L29, L31 | 10/10 | 8/10 |
| top_by_Cminus | L30, L31 | 10/10 | 9/10 |
| empos_contrast | L31 | 10/10 | 1/10 |
| mask_delta | L0, L4, L11, L13, L15, L21, L24, L26… | 10/10 | 0/10 |
| random | L1, L2, L8, L9, L13, L17, L22, L29 | 10/10 | 10/10 |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5382 | +1.143 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/prefixed words (e.g. "pre-", "multi-", "-to-", "-per-"), i.e. it ma… |
| 2 | L31·N1016 | -0.3720 | +0.778 | Fires on subword fragments in the middle of all-caps acronyms, product codes and brand names (e.g. "GR|AB", "NW|OT", "CPC|B", "15,000|th|b", "Cannon" … |
| 3 | L31·N7123 | +0.3093 | +0.700 | Mildly positive on the opening piece of a highly predictable multi-token span — a time-unit word right after a numeral ("1 year ago", "4 years, 11 mon… |
| 4 | L31·N10982 | +0.2965 | +1.073 | Fires at completion boundaries — the token that closes off a word or clause: final sub-word pieces of multi-token words (LET in CHEVRO|LET, ING in CHO… |
| 5 | L30·N10548 | +0.1013 | +0.757 | Fires on word-initial "B"-consonant subword fragments starting abbreviations/proper names (BR, Bh, bl, Br, and similar capital-letter fragments like "… |
| 6 | L31·N4172 | +0.0987 | +1.499 | Fires on sub-word fragments that complete rare compound/foreign/technical words (and on repeated tokens in duplicated list-like strings), e.g. the "ly… |
| 7 | L31·N7569 | +0.0850 | +0.882 | Fires on the opening delimiter (usually " (" , also an opening curly quote) that introduces a parenthesized acronym/abbreviation right after a fully s… |
| 8 | L31·N5053 | -0.0759 | +1.202 | Fires on tokens inside machine-ish identifier strings — URL/domain pieces, hardware IDs, product/statute codes, and comma-separated tag lists — while … |
| 9 | L31·N8597 | -0.0681 | +1.864 | Fires on the first piece of a multi-token compound proper noun / hyphenated or split term — the boundary point in things like "SALEM**,** ORE.", "**Wi… |
| 10 | L30·N5904 | +0.0636 | +0.952 | Fires on a token that opens a highly predictable fixed collocation/idiom — especially "way" in "(all the) way back", "the" in "the building blocks", "… |
| # | neuron | attr_sum | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5382 | +1.143 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/prefixed words (e.g. "pre-", "multi-", "-to-", "-per-"), i.e. it ma… |
| 2 | L31·N1016 | -0.3720 | +0.778 | Fires on subword fragments in the middle of all-caps acronyms, product codes and brand names (e.g. "GR|AB", "NW|OT", "CPC|B", "15,000|th|b", "Cannon" … |
| 3 | L31·N7123 | +0.3093 | +0.700 | Mildly positive on the opening piece of a highly predictable multi-token span — a time-unit word right after a numeral ("1 year ago", "4 years, 11 mon… |
| 4 | L31·N10982 | +0.2965 | +1.073 | Fires at completion boundaries — the token that closes off a word or clause: final sub-word pieces of multi-token words (LET in CHEVRO|LET, ING in CHO… |
| 5 | L30·N10548 | +0.1198 | +0.757 | Fires on word-initial "B"-consonant subword fragments starting abbreviations/proper names (BR, Bh, bl, Br, and similar capital-letter fragments like "… |
| 6 | L22·N5783 | -0.1094 | — | |
| 7 | L31·N4172 | +0.0987 | +1.499 | Fires on sub-word fragments that complete rare compound/foreign/technical words (and on repeated tokens in duplicated list-like strings), e.g. the "ly… |
| 8 | L13·N4272 | -0.0956 | — | |
| 9 | L28·N11638 | +0.0919 | +1.276 | Fires on the token immediately preceding a capital "B" (in enumerations "a)…and b)", formulas "A squared + B squared", color/abbrev lists "G: 48, B:",… |
| 10 | L31·N7569 | +0.0850 | +0.882 | Fires on the opening delimiter (usually " (" , also an opening curly quote) that introduces a parenthesized acronym/abbreviation right after a fully s… |
| # | neuron | attr_max | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | -0.5382 | +1.143 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/prefixed words (e.g. "pre-", "multi-", "-to-", "-per-"), i.e. it ma… |
| 2 | L31·N1016 | -0.3720 | +0.778 | Fires on subword fragments in the middle of all-caps acronyms, product codes and brand names (e.g. "GR|AB", "NW|OT", "CPC|B", "15,000|th|b", "Cannon" … |
| 3 | L31·N7123 | +0.3093 | +0.700 | Mildly positive on the opening piece of a highly predictable multi-token span — a time-unit word right after a numeral ("1 year ago", "4 years, 11 mon… |
| 4 | L31·N10982 | +0.2965 | +1.073 | Fires at completion boundaries — the token that closes off a word or clause: final sub-word pieces of multi-token words (LET in CHEVRO|LET, ING in CHO… |
| 5 | L30·N10548 | +0.1013 | +0.757 | Fires on word-initial "B"-consonant subword fragments starting abbreviations/proper names (BR, Bh, bl, Br, and similar capital-letter fragments like "… |
| 6 | L31·N4172 | +0.0987 | +1.499 | Fires on sub-word fragments that complete rare compound/foreign/technical words (and on repeated tokens in duplicated list-like strings), e.g. the "ly… |
| 7 | L22·N5783 | -0.0950 | — | |
| 8 | L31·N7569 | +0.0850 | +0.882 | Fires on the opening delimiter (usually " (" , also an opening curly quote) that introduces a parenthesized acronym/abbreviation right after a fully s… |
| 9 | L31·N5053 | -0.0759 | +1.202 | Fires on tokens inside machine-ish identifier strings — URL/domain pieces, hardware IDs, product/statute codes, and comma-separated tag lists — while … |
| 10 | L31·N8597 | -0.0681 | +1.864 | Fires on the first piece of a multi-token compound proper noun / hyphenated or split term — the boundary point in things like "SALEM**,** ORE.", "**Wi… |
| # | neuron | Cplus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N10982 | +0.2965 | +1.073 | Fires at completion boundaries — the token that closes off a word or clause: final sub-word pieces of multi-token words (LET in CHEVRO|LET, ING in CHO… |
| 2 | L31·N4172 | +0.0987 | +1.499 | Fires on sub-word fragments that complete rare compound/foreign/technical words (and on repeated tokens in duplicated list-like strings), e.g. the "ly… |
| 3 | L31·N7569 | +0.0850 | +0.882 | Fires on the opening delimiter (usually " (" , also an opening curly quote) that introduces a parenthesized acronym/abbreviation right after a fully s… |
| 4 | L31·N10854 | +0.0454 | +1.202 | Fires on a token that begins a highly predictable fixed collocation/compound — i.e., the word right before an almost-certain continuation ("re"+brandi… |
| 5 | L28·N11638 | +0.0437 | +1.276 | Fires on the token immediately preceding a capital "B" (in enumerations "a)…and b)", formulas "A squared + B squared", color/abbrev lists "G: 48, B:",… |
| 6 | L31·N4764 | +0.0361 | +1.374 | Fires on an opening quotation mark (curly “ / ‘ or straight ") when it immediately follows the indefinite article "a", i.e., at the start of a scare-q… |
| 7 | L31·N4987 | +0.0286 | +1.009 | Fires on the final content word that closes a noun/verb phrase (a head noun, adjective or past-participle verb completing a description, often in info… |
| 8 | L29·N8536 | +0.0283 | +1.271 | Fires in "label + letter/number" naming constructions (e.g. "Building B", "Group B", "Semi Final 2", "Addendum 4‑A", "stage 3"), peaking on the head n… |
| 9 | L21·N8004 | +0.0269 | +0.890 | Fires on the delimiter punctuation of the first enumerated list item ("1." / "1)") — i.e., the period or paren right after a leading "1" that opens a … |
| 10 | L19·N1211 | +0.0225 | +1.402 | Fires on the continuation token that completes a word split across sub-word pieces (City|search, Clo|sets, H|elen) or on filler tokens inside boilerpl… |
| # | neuron | Cminus | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N4902 | +0.5363 | +1.143 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/prefixed words (e.g. "pre-", "multi-", "-to-", "-per-"), i.e. it ma… |
| 2 | L31·N1016 | +0.3720 | +0.778 | Fires on subword fragments in the middle of all-caps acronyms, product codes and brand names (e.g. "GR|AB", "NW|OT", "CPC|B", "15,000|th|b", "Cannon" … |
| 3 | L31·N5053 | +0.0759 | +1.202 | Fires on tokens inside machine-ish identifier strings — URL/domain pieces, hardware IDs, product/statute codes, and comma-separated tag lists — while … |
| 4 | L31·N8597 | +0.0681 | +1.864 | Fires on the first piece of a multi-token compound proper noun / hyphenated or split term — the boundary point in things like "SALEM**,** ORE.", "**Wi… |
| 5 | L31·N9287 | +0.0611 | +1.312 | Fires on tokens inside runs of ALL-CAPS headings / Title-Case phrases (often the first fragment of a word split across tokens, e.g. "COM|E", "Logo|Bug… |
| 6 | L31·N8648 | +0.0468 | +1.264 | Fires on the final token of a name or number string that is *continued* by more text (e.g., last letter of "Tabitha" before the surname, the "8" of a … |
| 7 | L31·N1720 | +0.0424 | +1.223 | Fires on tokens inside runs of ALL-CAPS or Title-Case text (headline/slogan-style capitalized word sequences), peaking on a mid-phrase capitalized wor… |
| 8 | L30·N1827 | +0.0421 | +0.876 | Fires on the token that continues/completes a fixed multi‑word expression or title being echoed from earlier context (induction‑style completion of se… |
| 9 | L30·N10237 | +0.0403 | +1.131 | Fires on highly predictable continuation tokens that complete a multi-token phrase, name, date, or clause already established earlier in the passage (… |
| 10 | L31·N5228 | +0.0391 | +2.187 | Fires on the first sub-word fragment of an uncommon multi-token word (names, compounds, jargon) that is about to be completed by suffix pieces — e.g. … |
| # | neuron | empos_contrast | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N8693 | -5.4664 | — | |
| 2 | L31·N11306 | +5.3076 | — | |
| 3 | L31·N4918 | +4.7233 | — | |
| 4 | L31·N12266 | +4.4072 | — | |
| 5 | L31·N8769 | -4.3724 | — | |
| 6 | L31·N6951 | -4.3326 | — | |
| 7 | L31·N2753 | +3.9151 | — | |
| 8 | L31·N10897 | -3.8973 | — | |
| 9 | L31·N10263 | +3.6870 | — | |
| 10 | L31·N4257 | -3.6485 | — |
| # | neuron | mask_delta | M | auto-interp description |
|---|---|---|---|---|
| 1 | L4·N7153 | +2.6549 | -1.655 | Fires on the paragraph-break newline (or the sentence-final period itself) that ends a complete declarative sentence in web/how-to prose, i.e. the seg… |
| 2 | L24·N474 | +2.6158 | -1.616 | Fires in manufacturing/product-description prose, peaking on the linking word that introduces how something is made or what it's made of — chiefly "is… |
| 3 | L0·N1876 | +2.6036 | +3.604 | This neuron is mainly defined by its negative pole — it is strongly suppressed by "depth"/"deep" words ("in-depth", "deeper", "deeply") — while its po… |
| 4 | L15·N9418 | +2.5040 | -1.504 | Fires on the head noun that completes a short definite noun phrase after a preposition (e.g. "from the **employer**", "within the **game**", "on the *… |
| 5 | L13·N5521 | +2.4824 | -1.482 | Fires on the highly predictable word that completes a familiar collocation or idiom, typically in generalizing statements about quantities of people/o… |
| 6 | L27·N9758 | +2.4507 | +3.451 | Fires on the first/initial sub-word piece of a multi-token proper name or title being spelled out mid-sentence (species, band-song, company, event nam… |
| 7 | L26·N11553 | +2.4190 | -1.419 | Fires on the modifier/verb token immediately preceding a noun denoting schoolwork or practice exercises (assignments, homework, problems, practice, in… |
| 8 | L21·N10342 | +2.3751 | -1.375 | This neuron fires on function-word markers of something forthcoming or not-yet-realized — "to" in infinitival "yet to be deployed"/"details to follow"… |
| 9 | L29·N7754 | +2.3728 | -1.373 | Fires on the *first* sub-token fragment of rare multi-piece proper nouns — domain/brand names, surnames, technical terms (e.g. ".hom|estead", "Vue|Sof… |
| 10 | L11·N9064 | +2.3676 | -1.368 | Fires on the very first token of a document/text snippet (often a capitalized word or an opening quotation mark beginning the text), rather than on an… |
| # | 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 | L31·N4902 | -0.5382 | +1.143 | Fires on the token immediately preceding a hyphen in hyphenated compound modifiers/prefixed words (e.g. "pre-", "multi-", "-to-", "-per-"), i.e. it ma… |
| 2 | L31·N10982 | +0.2965 | +1.073 | Fires at completion boundaries — the token that closes off a word or clause: final sub-word pieces of multi-token words (LET in CHEVRO|LET, ING in CHO… |
| 3 | L31·N4172 | +0.0987 | +1.499 | Fires on sub-word fragments that complete rare compound/foreign/technical words (and on repeated tokens in duplicated list-like strings), e.g. the "ly… |
| 4 | L31·N5053 | -0.0759 | +1.202 | Fires on tokens inside machine-ish identifier strings — URL/domain pieces, hardware IDs, product/statute codes, and comma-separated tag lists — while … |
| 5 | L31·N8597 | -0.0681 | +1.864 | Fires on the first piece of a multi-token compound proper noun / hyphenated or split term — the boundary point in things like "SALEM**,** ORE.", "**Wi… |
| 6 | L31·N9287 | -0.0611 | +1.312 | Fires on tokens inside runs of ALL-CAPS headings / Title-Case phrases (often the first fragment of a word split across tokens, e.g. "COM|E", "Logo|Bug… |
| 7 | L30·N11231 | +0.0510 | +1.086 | Fires on the sentence-terminating period "." at the end of declarative statements (and is strongly suppressed by "?" ending questions, e.g. in Q&A for… |
| 8 | L31·N8648 | -0.0468 | +1.264 | Fires on the final token of a name or number string that is *continued* by more text (e.g., last letter of "Tabitha" before the surname, the "8" of a … |
| 9 | L31·N10854 | +0.0454 | +1.202 | Fires on a token that begins a highly predictable fixed collocation/compound — i.e., the word right before an almost-certain continuation ("re"+brandi… |
| 10 | L28·N11638 | +0.0437 | +1.276 | Fires on the token immediately preceding a capital "B" (in enumerations "a)…and b)", formulas "A squared + B squared", color/abbrev lists "G: 48, B:",… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L31·N1016 | -0.3720 | +0.778 | Fires on subword fragments in the middle of all-caps acronyms, product codes and brand names (e.g. "GR|AB", "NW|OT", "CPC|B", "15,000|th|b", "Cannon" … |
| 2 | L31·N7123 | +0.3093 | +0.700 | Mildly positive on the opening piece of a highly predictable multi-token span — a time-unit word right after a numeral ("1 year ago", "4 years, 11 mon… |
| 3 | L30·N10548 | +0.1013 | +0.757 | Fires on word-initial "B"-consonant subword fragments starting abbreviations/proper names (BR, Bh, bl, Br, and similar capital-letter fragments like "… |
| 4 | L31·N7569 | +0.0850 | +0.882 | Fires on the opening delimiter (usually " (" , also an opening curly quote) that introduces a parenthesized acronym/abbreviation right after a fully s… |
| 5 | L30·N1827 | -0.0421 | +0.876 | Fires on the token that continues/completes a fixed multi‑word expression or title being echoed from earlier context (induction‑style completion of se… |
| 6 | L21·N8004 | +0.0269 | +0.890 | Fires on the delimiter punctuation of the first enumerated list item ("1." / "1)") — i.e., the period or paren right after a leading "1" that opens a … |
| 7 | L31·N5030 | +0.0176 | +0.900 | Fires on the medial word of fixed multi-word proper names/compound terms (where the next word is highly predictable), e.g. "Outlook **Web** Apps", "Fr… |
none
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L11·N9064 | -0.0002 | -1.368 | Fires on the very first token of a document/text snippet (often a capitalized word or an opening quotation mark beginning the text), rather than on an… |
| 2 | L21·N10342 | +0.0002 | -1.375 | This neuron fires on function-word markers of something forthcoming or not-yet-realized — "to" in infinitival "yet to be deployed"/"details to follow"… |
| 3 | L24·N474 | +0.0002 | -1.616 | Fires in manufacturing/product-description prose, peaking on the linking word that introduces how something is made or what it's made of — chiefly "is… |
| 4 | L13·N5521 | +0.0001 | -1.482 | Fires on the highly predictable word that completes a familiar collocation or idiom, typically in generalizing statements about quantities of people/o… |
| 5 | L28·N1724 | -0.0001 | -1.344 | Fires on sub-word pieces inside foreign/exotic proper names and mojibake-encoded Spanish text (e.g. the "ñ" of "Piñata", the "R" of RFEF, "ash" of V… |
| 6 | L15·N9418 | -0.0001 | -1.504 | Fires on the head noun that completes a short definite noun phrase after a preposition (e.g. "from the **employer**", "within the **game**", "on the *… |
| 7 | L4·N7153 | +0.0000 | -1.655 | Fires on the paragraph-break newline (or the sentence-final period itself) that ends a complete declarative sentence in web/how-to prose, i.e. the seg… |
| 8 | L29·N7754 | -0.0000 | -1.373 | Fires on the *first* sub-token fragment of rare multi-piece proper nouns — domain/brand names, surnames, technical terms (e.g. ".hom|estead", "Vue|Sof… |
| 9 | L26·N11553 | +0.0000 | -1.419 | Fires on the modifier/verb token immediately preceding a noun denoting schoolwork or practice exercises (assignments, homework, problems, practice, in… |
| # | neuron | attr_last | M | auto-interp description |
|---|---|---|---|---|
| 1 | L30·N5904 | +0.0636 | +0.952 | Fires on a token that opens a highly predictable fixed collocation/idiom — especially "way" in "(all the) way back", "the" in "the building blocks", "… |
| 2 | L31·N7027 | -0.0342 | +1.000 | Fires on the preposition " in" when it introduces a place/location phrase (e.g., "a dentist **in** Camarillo, CA", "adoption **in** southeast Michigan… |
| 3 | L31·N4987 | +0.0286 | +1.009 | Fires on the final content word that closes a noun/verb phrase (a head noun, adjective or past-participle verb completing a description, often in info… |
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