jbloomAus / SAELens

Training Sparse Autoencoders on Language Models
https://jbloomaus.github.io/SAELens/
MIT License
193 stars 67 forks source link

[Bug Report] ActivationsStore fails for models without a tokenizer #186

Open chanind opened 2 weeks ago

chanind commented 2 weeks ago

Describe the bug For models without a tokenizer, e.g. Othello, an error is thrown in the ActivationsStore since it assumes a tokenizer is present and attempts to insert BOS tokens. If the model has no Tokenizer, we should gracefully handle this and not attempt to read anything form the tokenizer.

Code example See #177

Checklist