HypotheSAEs is a method which hypothesizes interpretable relationships in text datasets by training Sparse Autoencoders (SAEs) on foundation model representations.
Paper:
https://arxiv.org/abs/2502.04382Code:
https://github.com/rmovva/HypotheSAEs
This data repo contains all of the datasets used for the experiments in the paper.
In total, we use five datasets to evaluate HypotheSAEs.
We evaluate with two synthetic datasets, where the goal is to… See the full description on the dataset page:
https://huggingface.co/datasets/rmovva/HypotheSAEs.
