search

ai-safety-foundation / sparse_autoencoder

Sparse Autoencoder for Mechanistic Interpretability
https://ai-safety-foundation.github.io/sparse_autoencoder/
MIT License
180 stars 39 forks source link

Make demo work #47

Closed ArthurConmy closed 11 months ago

ArthurConmy commented 11 months ago

@alan-cooney what's up with the demo deleting the pipeline? Is there currently a workaround for the next two hours? Can I revert the deletions or will that cause problems?

ArthurConmy commented 11 months ago

FYI the demo currently uses a SoLU model (not GeLU), Pile data and GPT-2 tokenizer; seems inconsistent

alan-cooney commented 11 months ago

Been refactoring but it should be working at https://github.com/ai-safety-foundation/sparse_autoencoder/pull/50 (will merge shortly)