CrossCoder Through Time

Repo aims to train a CrossCoder model across multiple checkpoints of a 70M Pythia model, and see how it evolves over training.

• Adapted from crosscoder-model-diff-replication

Two ways to train the model:

1. On the fly: Cache activations as the model is trained
2. Cached: Cache activations ahead of time, then train the SAE on these cached activations.
   • Supports multi-gpu!
   • Takes up a lot of space on disk

• train.py: Entry point, define config and hyperparameters
• cache_acts.py: Cache activations ahead of time

Crosscoder lib:

• trainer.py: Simple training loop that can use either cached or on-the-fly activations
• model.py: Defines the CrossCoder model
• buffer_on_the_fly.py: Caches activations as the model is trained
• buffer_cached.py: Loads cached activations from disk