Aaquib111/edge-attribution-patching — reverse-engineered prompt

Build me a clean runnable version of this research project for edge attribution patching. I want to be able to open the repo, install the requirements, and run example experiments that show how attribution patching finds important edges in transformer circuits and compares against automated circuit discovery.

Please organize it so there are clear entry points for the IOI task, greater than task, docstring task, and threshold investigation, with simple notebooks or scripts that load the model, run the patching, save results, and make the existing drawings or plots easy to reproduce.

Keep it close to the paper, but make it friendlier for someone who is not deep in the code. Add a short README with setup steps, expected commands, what each experiment does, and any GPU assumptions. It is okay to use TransformerLens, and look up current docs online if needed.