SAELens — обучение разреженных автокодировщиков

★ 7.7 · ai-tooling

saelens is a Claude Code skill that guides training and analyzing Sparse Autoencoders (SAEs) using the SAELens library to decompose polysemantic neural network activations into sparse, interpretable features. Built on Anthropic's monosemanticity research — where human evaluators found 70% of SAE-discovered features genuinely interpretable, including DNA sequences, legal language, HTTP requests, and grammatical structures — the skill covers two main workflows: loading pre-trained SAEs (such as gpt2-small-res-jb or Gemma 2B releases) and training custom SAEs with configurable architecture (standard, gated, topk), sparsity coefficients, and expansion factors. Dependencies are sae-lens>=6.0.0, transformer-lens>=2.0.0, and torch>=2.0.0. It is aimed at researchers working on mechanistic interpretability, superposition analysis, feature-based steering or ablation, and safety-relevant feature discovery in language models.