Llama 3.2 Interpretability with Sparse Autoencoders

Open-source PyTorch pipeline for studying how Llama 3.2 works internally using sparse autoencoders. It is reproducible and aimed at interpretability researchers.