Lets anyone generate and explore attribution graphs on models such as Gemma-2-2b and Llama-3.2-1b, and test hypotheses by editing feature values. Interpretability work no longer requires a private stack.
Key takeaways · AI-distilled
Anthropic notes attribution graphs only partially reveal a model's internal steps; the Neuronpedia frontend lets users annotate and share graphs, not just view them.
Anthropic has already used the tools to study multi-step reasoning and multilingual representations in Gemma-2-2b and Llama-3.2-1b, and published unanalyzed graphs for the community to investigate.
The library was built by Anthropic Fellows with Decode Research, which built the Neuronpedia integration; the Gemma graphs rely on transcoders trained in the GemmaScope project.
