Decode Research

Fast, standardized, and easy-to-use interpretability engine.

Training Sparse Autoencoders on Language Models

code for the paper: SynthSAEBench: Evaluating Sparse Autoencoders on Scalable Realistic Synthetic Data

Code accompanying the paper "Are Sparse Autoencoder Benchmarks Reliable?"