event-sae-openpi-libero

Pre-trained BatchTopK SAE checkpoints for π₀.₅ (PaliGemma vision- language prefix + action expert) on LIBERO-Spatial. These are the post-MLP residual-stream SAEs from the "Event-Grounded Sparse Autoencoders for Vision-Language-Action Policies" project, ready to use with the Event-SAE intervention pipeline.

Files

Six checkpoints, one per (capture target, layer):

Capture target	Layer	Folder	Activation dim	Dict size
action_expert	0	action_expert_l00/	1024	1024
action_expert	5	action_expert_l05/	1024	1024
action_expert	17	action_expert_l17/	1024	1024
paligemma	0	paligemma_l00/	2048	2048
paligemma	11	paligemma_l11/	2048	2048
paligemma	16	paligemma_l16/	2048	2048

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

All six checkpoints share:

Hyperparameter	Value
Active budget k	64
Architecture	BatchTopKSAE
Hook location	post-block residual
Submodule	post_mlp_residual (AE) /
	post_mlp_residual__paligemma (PG)

Loading

from huggingface_hub import hf_hub_download
from dictionary_learning.trainers.batch_top_k import BatchTopKSAE

ckpt = hf_hub_download(
    "mr-cabbage/event-sae-openpi-libero", "action_expert_l17/ae.pt"
)
sae = BatchTopKSAE.from_pretrained(ckpt)

Code

Full pipeline (activation collection, SAE training, event clustering, feature ranking, residual-preserving intervention) is at github.com/xc-j/Event-SAE. The openpi half uses an openpi fork for the JAX io_callback SAE hook.

License

MIT.