Fluent dreaming for language models.

The code here implements the algorithms in the paper "Fluent dreaming for language models".

Please also see the companion page that demonstrates using the code here.

There is also a Colab version of the companion page.

Modules:

• dreamy.epo: The main EPO algorithm along with a few utilities for loading models and constructing Pareto frontiers.
• dreamy.attribution: Code for creating causal attribution figures.
• dreamy.runners: "Runners" for different optimization targets like neurons, output logits, etc.
• dreamy.experiment: Code we used in writing the paper for running experiments on Modal and using S3 for storage.