event-sae-openvla-libero

Pre-trained BatchTopK SAE checkpoints for openVLA on the four LIBERO simulation suites. These are the layer-31 post-MLP residual-stream SAEs used to produce paper Table 4 of "Event-Grounded Sparse Autoencoders for Vision-Language-Action Policies" (NIPS submission).

Files

Suite	Folder
LIBERO-Spatial	libero_spatial/
LIBERO-Object	libero_object/
LIBERO-Goal	libero_goal/
LIBERO-10	libero_10/

Each folder contains ae.pt (model weights) and config.json (trainer + dictionary config; required by the loader).

All four checkpoints share:

Hyperparameter	Value
Activation dim	4096
Dict size	32768
Active budget k	64
Hook location	post-block residual, layer 31
Architecture	BatchTopKSAE

Loading

from huggingface_hub import snapshot_download
from event_sae.openvla.activations import load_batch_topk_sae

local_dir = snapshot_download("mr-cabbage/event-sae-openvla-libero")
sae, config = load_batch_topk_sae(f"{local_dir}/libero_spatial/ae.pt", device="cuda")

Code

Full pipeline (activation collection, SAE training, event clustering, feature ranking, residual-preserving intervention) is at github.com/xc-j/Event-SAE.

License

MIT.