Plausibly, I should just precompute interp data for _all_ model layers, and commit those .csv files to GitHub

DavidUdell / sparse_circuit_discovery

Circuit discovery in GPT-2 small, using sparse autoencoding

MIT License

7 stars 1 forks source link

Plausibly, I should just precompute interp data for _all_ model layers, and commit those .csv files to GitHub #65

Closed DavidUdell closed 8 months ago

DavidUdell commented 8 months ago

This would greatly speed up my interp loop. Basically, I would comment contexts.py out of pipe.py, and carefully hold all the training hyperparameters constant while doing so.