Skip to the content.

SAE-feature operands per operator

The token-operand catalog says which tokens an operator binds; this says which monosemantic SAE features it reads, and whether its OV copies that content (+) or suppresses it (−) — the dossier’s section-G layer. READ = the attention-weighted dominant key-feature (the SAE feature most present where the head attends; content-filtered, glossed by top tokens). copy-score = the OV→unembed diagonal on that feature’s own tokens. Provisional, single corpus (Shakespeare prose); _ = leading space.

Only content / circuit operators bind on content; positional / addressing ops (prev-token, local, sink, self) attend by position, so their read-feature is incidental — the copy-score column is load-bearing.

gpt2 — jbloom/GPT2-Small-SAEs-Reformatted

operator head kind reads (SAE feature) copy-score (OV)
duplicate 0.5 content _you; _the; US +0.09 (copies)
self 0.1 addressing _you; US; _the +0.07 (copies)
structural 3.1 structural _your/Your/_Your; The; _it/it -0.03 (≈neutral)
prevtok 4.11 positional US; cius/ius; I +0.10 (copies)
local 4.11 positional US; cius/ius; I +0.10 (copies)
induction 5.5 content First; What/_What/_what; US +0.09 (copies)
sink 7.2 addressing US/us; _Citizen/_citizens; I +0.00 (≈neutral)
s_inhibition 7.3 circuit US/us; MAR/MEN/CI; _you/you/You +0.01 (≈neutral)
name_mover 9.6 circuit _Citizen/_citizens; MEN/_men/_Men; _it/’d/_are +0.04 (copies)
backup_name_mover 9.0 circuit _it/’d/_are; _Citizen/_citizens; US/us +0.03 (copies)
coreference 9.0 circuit _it/’d/_are; _Citizen/_citizens; US/us +0.03 (copies)
negative_mover 10.7 circuit ‘d/_not/_and; And/That/_and; _Citizen/_citizens -0.01 (≈neutral)

gemma-2-2b — Gemma Scope (gemma-scope-2b-pt-res, JumpReLU)

operator head kind reads (SAE feature) copy-score (OV)
duplicate 1.4 (SAE L0, head L1) content cius/▁belly/VIR; ▁the; UMN/GIL/▁Cai -0.15 (suppresses)
sink 0.3 addressing cius/▁belly/VIR; UMN/GIL/▁Cai; ▁Citizen +0.10 (copies)
local 0.0 positional cius/▁belly/VIR; UMN/GIL/▁Cai; ▁the +0.02 (≈neutral)
induction 6.3 content First; ▁gods/▁run/▁petition; Before +0.06 (copies)
prevtok 21.7 positional ▁Citizen/./▁belly; ▁the/▁a/▁to; First/▁first -0.10 (suppresses)
self 25.7 (SAE L24, head L25) addressing ▁the/⏎⏎/.; First; ▁the/▁your/▁own -0.21 (suppresses)
structural 24.6 structural ▁the/⏎⏎/.; First; ▁with/With/▁With +0.11 (copies)

GPT-2 has all-layer SAEs (exact per-head layer); Gemma Scope is 8 layers so each Gemma op uses its nearest available SAE layer (offset ≤1, annotated). Gemma’s read-features come out noisier than GPT-2’s — its heads put heavy attention on <bos>/structural tokens on this non-repetitive prose, so the dominant *content* key-feature is weaker (a corpus + attention-budget effect, not a tooling one); the copy-score still uses the head’s exact OV. The cached Qwen SAE is for qwen2-0.5b (a different model). Data: operator_sae_operands_summary.json. Regenerate: operator_sae_operands.py.