Add NNsight skills benchmark suite #1
Draft
Conversation
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
Add a comprehensive benchmark suite for evaluating NNsight skills across different difficulty levels (easy, medium, hard) and techniques (basics, logit-lens, activation-patching, attribution-patching, causal-tracing, model-steering). The benchmark includes: - 17 query files testing specific interpretability techniques - Schema definitions for queries and results - Structural validation for API correctness - Deprecated pattern detection (pre-0.5 NNsight) - Claude Code and API runners - Results analysis and comparison tools Filenames now use descriptive hyphen-separated titles based on the query content rather than generic numeric IDs.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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
This PR adds a comprehensive benchmark suite for evaluating NNsight skills across different difficulty levels and interpretability techniques.
What's included
17 query files organized by difficulty (6 easy, 6 medium, 5 hard) covering:
nnsight-basics: Core tracing, saving, interventionslogit-lens: Layer-wise prediction decodingactivation-patching: Causal intervention via swappingattribution-patching: Gradient-based approximationcausal-tracing: Mediation analysismodel-steering: Steering vectors, persistent editsInfrastructure:
schema.py)validators/structural.py)validators/deprecated.py)runners/)analyze.py)Documentation:
Filename changes
Query files now use descriptive hyphen-separated titles based on their content (e.g.,
extract-hidden-states.yaml,position-specific-head-patching.yaml) instead of generic numeric IDs.Test plan
python skills_benchmark/runners/base.py --num-runs 1