Sparse Autoencoder for Mechanistic Interpretability

Work on sparse coding, replicating and extending the sparse coding approach to taking transformer features out of superposition.

Sparse probing paper full code.

Keeping language models honest by directly eliciting knowledge encoded in their activations.

A library for mechanistic interpretability of GPT-style language models

🤗 PEFT: State-of-the-art Parameter-Efficient Fine-Tuning.

Multiversal tree writing interface for human-AI collaboration