We are a group of hackers from Stanford's NLP group, and we are interested in LLM interpretability.
pyvene
is where we started, which stands for pytorch model intervenetion.
Supervised dictionary learning models (SDLs) and datasets releases for Gemma 2 2B and 9B: AxBench Collection
.
Benchmark interpretability methods at scale (AxBench) library: AxBench
.
Representation finetuning (ReFT) library: pyreft
.
PyTorch model intervention library: pyvene
.