qwen35_l12_oft_libero10_40k

Trained with starVLA-exp (run qwen35_l12_oft_libero10_20261012_062531).

Framework	QwenOFT
Base VLM	playground/Pretrained_models/Qwen3.5-0.8B
select_layer (LLM layers kept)	12
Action horizon	8
Training data	libero_10 (lerobot_datasets)
Images	agentview + wrist, 224 px; state: False
Optimizer steps	50000
Global batch	16
Learning rates	{'action_model': 0.0001, 'base': 2.5e-05, 'qwen_vl_interface': 1e-05}
Warmup / schedule	3000 steps / cosine_with_min_lr
Training energy	1.150609 kWh on NVIDIA GeForce RTX 5090
starVLA-exp commit	c13c4c450bbfd12e15025bf77d9642ca631702d9

Checkpoints in this repo:

• checkpoints/steps_40000_pytorch_model.pt

Use

The repo is a starVLA run directory (config.yaml + dataset_statistics.json + checkpoints/). It needs the starVLA-exp code (layer truncation, per-tokenizer OFT action token), not upstream starVLA.

# vla-evaluation-harness (starexp server), LIBERO-Long; downloads only the chosen checkpoint:
vla-eval serve -c configs/model_servers/starexp/libero.yaml \
    --arg checkpoint=Zenma/qwen35_l12_oft_libero10_40k --arg checkpoint_step=<N>
# every checkpoint of this repo:
bash scripts/sweep_starexp.sh --hf-repo Zenma/qwen35_l12_oft_libero10_40k
# or a plain local copy (a regular run directory):
hf download Zenma/qwen35_l12_oft_libero10_40k --local-dir qwen35_l12_oft_libero10_40k