Downloads · 30 days
0
starvla-grpo-libero-spatial-20260901
HOJQVFNA/starvla-grpo-libero-spatial-20260901
starvla-grpo-libero-spatial-20260901 is a reinforcement learning model from HOJQVFNA. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. It is set up for pytorch. The card lists the license as apache-2.0.
Full-parameter GRPO fine-tuning of StarVLA / Qwen2.5-VL-OFT on LIBERO-Spatial, evaluated with a SimpleVLA-RL-aligned protocol.
Access
Public
Updated Sep 2, 2026
Repo size
24.6 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt24.6 GB · 100%
At a glance
- Task
- Reinforcement Learning
- Library
- pytorch
- License
- apache-2.0
- Access
- Public
- Created
- Sep 2, 2026
- Updated
- Sep 2, 2026
- SHA
- dd2238ca
Datasets
Model card
From the Hugging Face model README
StarVLA GRPO on LIBERO-Spatial (RLinf)
Full-parameter GRPO fine-tuning of StarVLA / Qwen2.5-VL-OFT on LIBERO-Spatial, evaluated with a SimpleVLA-RL-aligned protocol.
Hardware
- Host: ASUS ESC8000A-E11 (
cscg103) - 2× AMD EPYC 7543 (64C / 128T), ~1 TiB RAM
- 8× NVIDIA GeForce RTX 4090 (450W), Driver 580.159.03
Official eval protocol
- 500 trajectories × 512 max steps
- Greedy decoding
- Seeds 0 / 1 / 2 → report mean ± std of
eval/success_once - Training-time
env/*metrics are optimistic and not the primary claim
Results (official)
| global_step | mean success | std |
|---|---|---|
| 0 (SFT) | 92.73% | 0.41% |
| 20 | 94.00% | 0.43% |
| 30 | 94.20% | 0.75% |
| 40 | 94.53% | 0.25% |
| 50 | 95.47% | 0.41% |
| 60 | 96.07% | 0.66% |
| 70 (best) | 96.67% | 0.52% |
| 80 | 96.27% | 0.68% |
| 90 | 96.47% | 0.66% |
Full table: results/official_eval_simplevla_rl.tsv
Repo layout
weights/
sft_baseline/steps_30000_pytorch_model.pt # init / step 0
grpo_global_step_70/full_weights.pt # best official
grpo_global_step_90/full_weights.pt # latest saved
results/ # TSV + per-ckpt eval_summary.json
eval_logs/full500_eval_campaign/ # seed-level eval logs
logs/ # train metrics + clean TB
Notes
- Intermediate GRPO steps 40/50/60/80 and DCP resume shards (
*.distcp, ~22G/step) are not included; ask if you need them. - Step 10 checkpoint was never saved.
- SFT init path on the training machine:
Qwen2.5-VL-OFT-LIBERO-4in1(steps_30000).
- Task
- Reinforcement Learning
- Library
- pytorch
- License
- apache-2.0
- Created
- Sep 2, 2026
- Updated
- Sep 2, 2026