docs: example — Cross-SAE alignment via FeatureMatch (cosine + top-k)

[BentleyRolling]

Summary

Adds an example notebook demonstrating cross-SAE feature alignment using the external FeatureMatch package (MIT, Course Correct Labs). Provides cosine similarity heatmap, top-k matches, and quick stats.

Why

Researchers often need to compare whether two SAEs learned similar dictionaries across seeds/hparams. This example standardizes a quick, reproducible check.

How to use

# !pip install "git+https://github.com/Course-Correct-Labs/featurematch.git"
from featurematch.featurematch import align_features
res = align_features(Z_a, Z_b, topk=5, threshold=0.8)
print(res.stats)

Notes

• Default cells run on synthetic data (works in CI/docs).
• A commented section shows how to collect real codes from SAELens SAEs (same hook/layer + dataset).
• Happy to upstream a minimal internal copy later if maintainers prefer tighter integration.

Checklist

• Example runs end-to-end on CPU
• No external keys or data required
• Clear interpretation guidance included