Skip to content

Example: add non-gradient attribution baselines - #366

Open
luciaquirke wants to merge 1 commit into
mainfrom
bank-baselines
Open

Example: add non-gradient attribution baselines#366
luciaquirke wants to merge 1 commit into
mainfrom
bank-baselines

Conversation

@luciaquirke

@luciaquirke luciaquirke commented Jul 28, 2026

Copy link
Copy Markdown
Collaborator

Two content-similarity LDS baselines:

  • Activation similarity over the attribution target modules.
  • Text semantic similarity with jina-embeddings-v3.

On the GPT-2 WikiText fine-tune re-train bank nicknamed lotus the mean LDSs are 0.129 and 0.095.

@luciaquirke luciaquirke changed the title Add non-gradient attribution baselines Example: add non-gradient attribution baselines Jul 28, 2026
@luciaquirke
luciaquirke force-pushed the bank-baselines branch 13 times, most recently from 619689e to 5b23e2c Compare July 29, 2026 12:45
Two content-similarity baselines scored against a re-train bank and evaluated by
per-query LDS, as a floor for the gradient methods:

- activation similarity over the attribution target modules (uses the bank
  model's activations)
- semantic search with jina-embeddings-v3 (pure text similarity)

Point either at any bank written with save_retrained_models=true via --bank; the
model and dataset are read from the bank's config.yaml, and LDS reuses
evaluate_retrained's cached per-subset losses. Omit --bank to build the default
GPT-2/WikiText bank first. On that bank they score mean per-query Spearman 0.13
and 0.09, below TrackStar (0.18), SOURCE (~0.39) and MAGIC (~0.92).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant