metadata
license: mit
datasets:
- monology/pile-uncopyrighted
library_name: transformers
This contains the weights of a sparse autoencoder I trained on the residual activations of Mistral-7B-Instruct-v0.1. I used The Pile (uncopyrighted) for the training data. As of right now, I have only trained a single SAE (on layer 16), though I may do more in the future.
Here is the associated GitHub repo I used to train it, which you can also use to do some crude inference.