Jiahao-Wang/lingbot-va-robotwin-2task-to-8task-full-step10000

LingBot-VA RoboTwin 2-Task-to-8-Task Full Fine-Tune Step 10000

This is a LingBot-VA inference bundle trained in two stages:

1. Full post-training on RoboTwin 2.0 blocks_ranking_rgb and stamp_seal for 5000 steps.
2. Continued full post-training on eight RoboTwin 2.0 tasks for 10000 steps.

The eight-task continuation used the default full post-training method, not LoRA.

Training tasks for the continuation stage:

• click_alarmclock
• stack_blocks_two
• handover_block
• place_object_basket
• pick_dual_bottles
• shake_bottle
• turn_switch
• beat_block_hammer

The bundle contains tokenizer/, text_encoder/, vae/, and the final merged transformer/. It can be used directly with wan_va_server.py --pretrained-model-path.