Instructions to use HuggingFaceVLA/smolvla_libero with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LeRobot
How to use HuggingFaceVLA/smolvla_libero with LeRobot:
# See https://github.com/huggingface/lerobot?tab=readme-ov-file#installation for more details git clone https://github.com/huggingface/lerobot.git cd lerobot pip install -e .[smolvla]
# Launch finetuning on your dataset python lerobot/scripts/train.py \ --policy.path=HuggingFaceVLA/smolvla_libero \ --dataset.repo_id=lerobot/svla_so101_pickplace \ --batch_size=64 \ --steps=20000 \ --output_dir=outputs/train/my_smolvla \ --job_name=my_smolvla_training \ --policy.device=cuda \ --wandb.enable=true
# Run the policy using the record function python -m lerobot.record \ --robot.type=so101_follower \ --robot.port=/dev/ttyACM0 \ # <- Use your port --robot.id=my_blue_follower_arm \ # <- Use your robot id --robot.cameras="{ front: {type: opencv, index_or_path: 8, width: 640, height: 480, fps: 30}}" \ # <- Use your cameras --dataset.single_task="Grasp a lego block and put it in the bin." \ # <- Use the same task description you used in your dataset recording --dataset.repo_id=HF_USER/dataset_name \ # <- This will be the dataset name on HF Hub --dataset.episode_time_s=50 \ --dataset.num_episodes=10 \ --policy.path=HuggingFaceVLA/smolvla_libero - Notebooks
- Google Colab
- Kaggle
Add Provael community eval results (.eval_results/provael.yaml)
Hi, this adds .eval_results/provael.yaml for this checkpoint: community eval entries from Provael, an open-source red-team harness for VLA policies (Apache-2.0).
What is in it: 11 entries on the LIBERO-Object suite, 10 tasks. The adversarial arms and the benign floor at 5 seeds each (400 episodes), plus four harmless-variation controls at 3 seeds. Every value is the fraction of episodes that left a keep-out envelope, with the counts, the 95% Wilson interval, the predicate state and the tool version in the notes, so the floor sits next to each rate. The headline: roleplay 42/50 against a benign floor of 1/50, and the same frame with no target named 27/30, so it reads as fragility under a long out-of-distribution imperative rather than steering. Clean task success on the benign arm was 48/50.
Benchmark dataset: Sattyam/provael-libero-object-redteam (eval.yaml). Each entry links the committed run directory with the per-task report.json files and execution manifests, so anyone can re-run it. Simulation only, uncalibrated default predicate, and none of it is HF Jobs verified; it shows as community-provided, which is the right label.
If you would rather not carry third-party eval files in this repo, say so and I will close this. Happy to adjust the format too.