sae-lens

Domain sparse autoencoders (SAELens) trained on LLM activations during fact-checking verdicts to find, validate and steer the latent behind abstention ("Not Enough Evidence").

A monorepo of AI safety experiments: sparse autoencoders on gelu-2l with activation steering, explainable AI on Kickstarter data using SHAP and LIME, and activation oracles on Qwen2.5-0.5B.

SAE Lens v6-compatible TDA notebook for mechanistic interpretability

Mechanistic Workbench (mwb): Local-first mechanistic interpretability workbench for IPython research, agent-readable state, artifact validation, evidence graphs, claim-safe MechanismCards, provenance, run ledgers, SAE/TransformerLens workflows, and reproducible MI experiments