jasper0314/starvla-libero-oft-2b

StarVLA QwenOFT 2B — LIBERO

This repository contains a StarVLA QwenOFT checkpoint fine-tuned on the combined LIBERO training set (libero_all). The vision-language backbone is Qwen/Qwen3-VL-2B-Instruct.

Model

Field	Value
Framework	StarVLA / QwenOFT
Backbone	Qwen3-VL-2B-Instruct
Action head	MLP, hidden size 2048
Action space	7D delta end-effector
Action horizon	8
Observation	One 224×224 RGB image plus language
Robot state	Not used
Training steps	30,000

The checkpoint is stored at:

final_model/pytorch_model.pt

Keep the repository layout intact. StarVLA resolves config.yaml and dataset_statistics.json relative to the checkpoint path.

LIBERO evaluation

Evaluation used 50 episodes per task, 500 episodes per suite, and 2,000 episodes in total.

Suite	Successes	Success rate
LIBERO-Object	496 / 500	99.2%
LIBERO-Goal	490 / 500	98.0%
LIBERO-Spatial	488 / 500	97.6%
LIBERO-10	434 / 500	86.8%
Overall	1,908 / 2,000	95.4%

The per-episode aggregate outputs are included under eval/. Evaluation used the synchronous vla-eval harness (0.3.1.dev18+g5a61b41f5.d20260712) with seed 7.

Download

hf download jasper0314/starvla-libero-oft-2b \
  --local-dir /path/to/starvla-libero-oft-2b

Pass the following checkpoint path to the StarVLA loader:

/path/to/starvla-libero-oft-2b/final_model/pytorch_model.pt

The portable config refers to the backbone by its Hugging Face repository ID, so the base model must also be available in the local Hugging Face cache or downloadable at model initialization time.

Reproducibility and licenses

The checkpoint was produced with StarVLA source commit 97af1dd41b366331939d1942294430c19501ab03. Full training settings are retained in config.full.yaml.

The Qwen3-VL base weights are distributed under Apache-2.0. StarVLA source code is distributed under its MIT license. Users are responsible for following the LIBERO dataset and benchmark terms.