jiyeony

Event-SAE Head full10 completed experiment

A public release of 3600 completed rollouts from a LIBERO-Spatial interpretability experiment using sparse autoencoders (SAE) with OpenVLA, including 18 conditions, 200 paired evaluation cases, and 16 ablated features for analyzing model behavior through feature interventions.

Downloads37
Episodes3600
Hours40.877

Why This Matters for Physical AI

This dataset enables research into neural network interpretability for robotic control by providing systematic ablation studies of learned features, helping understand which model components are critical for task success in embodied AI systems.

Technical Profile

Environment
simulation
Episodes
3600
Total Hours
40.877
Data Format
tar.gz
Annotation Types
language_instructionsreward_labelsaction_labels
Part of the LIBERO family

Access

Need custom physical AI data?

Claru builds purpose-built datasets for simulation applications with dense human annotations and quality assurance.

Request a Sample Pack

Related Datasets