Expert10k

LingBot-VLA v2 6B post-trained on RoboTwin with flow-matching behavior cloning.

• Training: all Action Expert layers and projections (1.789B parameters); vision-language backbone frozen.
• Data: clean demonstrations from 50 tasks; 10,000 updates, global batch 64, no gradient accumulation.
• Weights: full BF16 checkpoint, 25 safetensors shards, approximately 12.77 GB; action chunk length 50.

The complete clean + randomized evaluation is pending.

After downloading, run python prepare_run.py to create the checkpoint layout required by the companion lingbot-vla-robotwin policy server. Normalization and robot/processor configurations are included.