mr-cabbage/event-sae-libero

Event-SAE: LIBERO checkpoints

Pretrained BatchTopK SAE checkpoints for OpenVLA and OpenPI (pi0.5) on LIBERO, accompanying Event-Grounded Sparse Autoencoders for Vision-Language-Action Policies (NeurIPS 2026).

Checkpoints

Each folder contains ae.pt and its matching config.json. All checkpoints use the transformer block output (resid_post), with active budget 64.

OpenVLA

Layer 31, activation dimension 4,096, dictionary size 32,768. Folders below are under openvla/.

Suite	Folder	Learning rate	Training steps	Version
LIBERO-Spatial	libero_spatial_lr5e-5/	5e-5	4,000	Paper checkpoint
LIBERO-Spatial	libero_spatial_lr1e-4/	1e-4	15,000	Retrained checkpoint
LIBERO-Object	libero_object/	1e-4	15,000	Paper checkpoint
LIBERO-Goal	libero_goal/	1e-4	15,000	Paper checkpoint
LIBERO-10	libero_10/	1e-4	15,000	Paper checkpoint

Use openvla/libero_spatial_lr5e-5/ for the paper's Spatial results and feature IDs. The lr1e-4 SAE was retrained with batch size 40,000; its feature IDs and rankings differ.

OpenPI (pi0.5)

Trained on LIBERO-Spatial. Folders below are under openpi/libero_spatial/.

Component	Layer	Folder	Activation dimension	Dictionary size
Action expert	0	action_expert_l00/	1,024	1,024
Action expert	5	action_expert_l05/	1,024	1,024
Action expert	17	action_expert_l17/	1,024	1,024
PaliGemma	0	paligemma_l00/	2,048	2,048
PaliGemma	11	paligemma_l11/	2,048	2,048
PaliGemma	16	paligemma_l16/	2,048	2,048

Loading

Install Event-SAE, then download a checkpoint with its configuration:

from huggingface_hub import snapshot_download
from event_sae.sae import load_batch_topk_sae

folder = "openvla/libero_spatial_lr5e-5"
local_dir = snapshot_download(
    "mr-cabbage/event-sae-libero",
    allow_patterns=[f"{folder}/ae.pt", f"{folder}/config.json"],
)
sae, config = load_batch_topk_sae(f"{local_dir}/{folder}/ae.pt", device="cuda")

For OpenPI, set folder = "openpi/libero_spatial/action_expert_l05" or choose another checkpoint from the tables above.

Code and license

The collection, training, event analysis, and intervention pipeline is available at xc-j/Event-SAE. MIT license.