interp-gpt2-onnx

Segmented ONNX export of GPT-2 (small) for in-browser mechanistic interpretability. The model is split into 14 graphs — embed, twelve transformer blocks (block_00–block_11) that also emit attention patterns, and unembed for the logit lens — so a static web app can chain them entirely client-side with onnxruntime-web (WebGPU/WASM). Provided in three precisions: fp32/, fp16/, and int8/, each with a manifest.json (byte sizes + content_hash for cache versioning).

Live demo: https://interp.zackwithers.com · Source: https://github.com/legendaryzoac/interp