Train, load, and analyze sparse autoencoders for language model interpretability.
npx skills add https://github.com/zjunlp/Mechanist --skill saelens-zjunlp
Please help me install this Agent Skill. Skill: saelens Source: https://github.com/zjunlp/Mechanist/tree/main/skills/mechanism-skills/feature-dictionary-learning/SAE Command: npx skills add https://github.com/zjunlp/Mechanist --skill saelens-zjunlp
High-intent search queries and answers about installing and using this skill.
Use the sae-lens package to run a cache activation pipeline that collects model activations, then train a Standard, TopK, or Gated SAE on them. The training tutorials in SAELens walk through configuration and evaluation of the trained SAE.
SAELens provides loading utilities that fetch pre-trained SAEs directly from Neuronpedia or HuggingFace by name. The pre-trained SAEs list in the documentation catalogs available releases such as Gemma Scope.
Yes, SAELens integrates with HuggingFace Transformers and other PyTorch-based models, and also provides HookedSAETransformer for hooking SAEs into TransformerLens forward passes during interpretability analysis.
Evaluate SAEs using variance explained, L0 sparsity, and reconstruction loss metrics built into SAELens. For standardized comparisons, the SAEBench benchmark suite provides additional evaluation protocols.
SAEs approximate activations with reconstruction error, so some information is lost, and learned features may not perfectly align with human-interpretable concepts. Results should be validated with metrics like reconstruction loss and cross-checked against downstream behavior.
jleechanorgTrain and analyze Sparse Autoencoders with SAELens to decompose neural activations.
math-incTrain and analyze Sparse Autoencoders on transformer activations with SAELens.
Monjyu1101Train and analyze Sparse Autoencoders on transformer activations using SAELens workflows.
0h-n0Sets up TransformerLens and nnsight intervention experiments for mechanistic interpretability analysis of LLMs.
