Breaking change
The package is renamed from oracle_sae to interp_lab. import oracle_sae no longer works — use import interp_lab. The interp-lab CLI and all command names are unchanged.
New
- Semantic text matching (opt-in). Text fingerprints can now use real semantic embeddings instead of lexical token hashing. Install
pip install "interp-lab[embeddings]"and pass--text-embedder minilm(local sentence-transformers MiniLM, free/offline) or setINTERP_LAB_TEXT_EMBEDDER. The default stays the dependency-free lexical hash. Each fingerprint records the embedder that produced it, and matching refuses to compare vectors from different embedders. - Archived real-model demo at
examples/real_model_demos/golden-distilgpt2-unit/— real DistilGPT-2 SAE artifacts with a measured criterion-promoting latent, an authentic suppression dose-response, and semantic (MiniLM) fingerprints.
Docs
- README tightened; the full command catalog moved to
docs/COMMANDS.md, which documents--text-embedder.
Tested: 211 passed, 1 skipped (the MiniLM test runs wherever the [embeddings] extra is installed).