Results
This section provides an overview of the main results obtained from the project. Two models were trained:
- JumpReLU SAE trained on cs.LG dataset.
- JumpReLU SAE trained on arxiv-astro-abstracts-all dataset.
The figure below shows the feature density histograms of both models.
For more details on the training results of the cs.LG SAE, see this W&B run. For more details on the training results of the arxiv-astro-abstracts-all SAE, see this other W&B run.