feat(atp-star): add QK correction, GradDrop, exact verification skill - #8
Closed
janmenjayap wants to merge 1 commit into
Closed
janmenjayap wants to merge 1 commit into
janmenjayap wants to merge 1 commit into
Conversation
Ports the agent-facing subset of the nnsight-side AtP* pattern (ndif-team/nnsight#705): exact local-softmax recompute for a saturated query/key, GradDrop's repeated-backward cancellation check, and the exact-verification workflow. Every example is real GPT-2, transformers>=5, TransformersModel. Cross-linked from attribution-patching in both directions. Deliberately left out (see PR description): the subset/confidence diagnostic (Student-t CDF, Algorithm 1 online stats) and the MVP benchmark protocol -- those are benchmark artifacts for a separate nnsight-side PR, not agent-facing failure-mode fixes.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
plugins/nnsight/skills/atp-star/SKILL.md, a new sibling skill next toactivation-patchingandcircuit-discovery, porting the agent-facing subset of thennsight-side AtP* pattern (feat(atp-star): add tested component-localization building blocks nnsight#705): exact local-softmax recompute for a
saturated Q/K pair, GradDrop's repeated-backward cancellation check, and the
exact-verification workflow.
0.8API directly (TransformersModel, transformers>=5) againstreal checkpoints (GPT-2), no synthetic tiny-random-weight models and no pre-0.8 idioms.
attribution-patching/SKILL.md's "Related skills" list nowpoints at
atp-star("an attribution-patching score looks near zero for a component yoususpect matters"), and
atp-starpoints back atattribution-patching(the baseline itcorrects) and
circuit-discovery.atp-starrow to the rootREADME.mdskill table and the.codex/skills/atp-starsymlink, matching the repo's existing per-skill layout.
Deliberately left out: the subset/confidence diagnostic (Student-t CDF, Algorithm 1 online
stats) and the MVP benchmark protocol — those are benchmark artifacts for the nnsight-side PR,
not agent-facing failure-mode fixes.
Test plan
make test-skill SKILL=atp-star(runs every fenced python block inSKILL.mdagainstreal models per this repo's
CLAUDE.md)make test-structure