metadata
license: mit
tags:
- sparse-autoencoder
- interpretability
- graphcast
- weather
- climate
GraphCast Climate SAEs
Sparse autoencoders trained on GraphCast layer activations for the Climate XAI project.
Six models: {Standard TopK, Matryoshka TopK} × layers {0, 8, 15}. Input dim 512,
4096 latents, k=32 active. Each directory holds final_model.pt and config.json.
Load with the code at https://github.com/mathaim/climate_xai
(train/models/plain_sae.py, train/models/matryoshka_sae.py); hyperparameters are in each config.json.