nanda-replication-codebase

Summary

This pins down a specific, publicly available starting point for Sparse Autoencoder work: a GitHub repo from Neel Nanda that reproduces SAE features on a deliberately simple one-layer GELU model, so anyone in the system can point to concrete code rather than just a paper. The fact that it uses a minimal base model means the learned features are easier to interpret and the setup is lightweight enough to rerun or modify without heavy compute.