This is an implementation of tied crosscoders, a variation of crosscoders especially useful for understanding how specific chat behavior arises from the base model. This is based on ckkissane's implementation of crosscoders.
- For playing with an already trained crosscoder, you can go to
analyze.ipynb.You will have to download the crosscoder from HuggingFace - For training a tied crosscoder, you should run
train.ipynb - For the visualizations inside this project, you should use this modified version of sae-vis.
- For convenience, I included 1M tokens from lmsys and 1M tokens from the pile in the datasetes folder.