LingBot-VLA action module — lora, step 10000
Inference release; existing repository visibility is preserved. Step-specific tags preserve all published versions. This snapshot has not been evaluated in simulation and is not a selected best model. Source code commit: 61bd361047c5c012ba39bbe3ad289f51b186194e. Details: release.json and metadata/.
Download and load
Download the complete snapshot with revision=step-10000 (or pin its full commit SHA). Keep the run layout: lingbotvla_cli.yaml + checkpoints/global_step_10000/hf_ckpt. From the downloaded snapshot root, use the existing LingBot-VLA v2 deployment code; set QWEN3VL_PATH=models/Qwen3-VL-4B-Instruct and the checkpoint argument to checkpoints/global_step_10000/hf_ckpt. Use BF16 inference and preserve action_fp32=true. No base Qwen weights are needed in the included tokenizer directory. Repository code and compatible simulation environment must be installed separately.
Original training config and source identity are in metadata/. Deployment config paths are snapshot-relative; run deployment with this snapshot as working directory. Original training datasets, optimizer state and DCP recovery shards are excluded. SHA256SUMS and file-manifest.json cover the release files.