Train and analyze sparse autoencoders for mechanistic interpretability of neural networks.
npx skills add https://github.com/NousResearch/hermes-agent --skill saelens
Please help me install this Agent Skill. Skill: saelens Source: https://github.com/NousResearch/hermes-agent/tree/main/optional-skills/mlops/saelens Command: npx skills add https://github.com/NousResearch/hermes-agent --skill saelens
Neural network neurons are polysemantic, activating across unrelated contexts due to superposition, which makes understanding what models have learned extremely difficult. This Skill provides a complete workflow for training Sparse Autoencoders (SAEs) that decompose dense activations into sparse, monosemantic features corresponding to interpretable concepts.
Use the saelens skill to load the gpt2-small-res-jb SAE and identify the top 10 features activating on the prompt 'The capital of France is'.
High-intent search queries and answers about installing and using this skill.
Use LanguageModelSAERunnerConfig with a nested SAE sub-config like StandardTrainingSAEConfig specifying d_in, d_sae, and l1_coefficient. Then instantiate LanguageModelSAETrainingRunner and call run() to train on activations collected from a specified hook point.
TransformerLens provides direct access to model activations and hooks for causal interventions. SAELens specifically trains sparse autoencoders to decompose those activations into interpretable features, solving the polysemanticity problem that raw neuron analysis cannot address.
Yes, SAELens v6 supports TopKTrainingSAEConfig where you set k directly on the SAE sub-config to enforce exactly K active features per token, providing consistent sparsity without L1 penalty tuning.
Dead features occur when L1 penalty is applied too aggressively early in training. Increase l1_warm_up_steps to gradually ramp up the sparsity penalty, or use ghost gradients to revive dead features during training.
Use SAELens when you need to discover and analyze interpretable features in model activations. Use pyvene when you need causal intervention experiments with predefined concepts, as pyvene focuses on activation patching rather than feature discovery.
AevonixTrain and analyze sparse autoencoders on Transformer model activations.
handsomelong922Train and analyze Sparse Autoencoders with SAELens to decompose neural activations.
zerwizTrain and analyze sparse autoencoders on transformer activations with SAELens.
hhhi21gPerform causal interventions on PyTorch model activations.
