ai-safety-foundation / sparse_autoencoder

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

[W/ Potential Severe Consequences] Running `sweep()` on different directory structure results in bugs. #208

Open HuFY-dev opened 5 months ago

HuFY-dev commented 5 months ago

I wrote a notebook calling the sweep function. sweep() calls wandb.agent(sweep_id, train) which calls train() which calls run_training_pipeline(), and in the release version of run_training_pipeline()#L268: checkpoint_path = Path("../../.checkpoints"), and this is used later to attempt to create a .checkpoint directory two directories up from my current working directory, which is messing up with my file system in the root directories. In the main branch, the code looks pretty different, which is instead creating .checkpoint with respect to where the library is installed (correct me if I'm wrong). I assume that's not expected behavior.