Jiahao-Wang/lingbot-va-robotwin-8task-lora-v2-r64-step10000

LingBot-VA RoboTwin 8-Task LoRA v2 r64 Step 10000

This is a LingBot-VA inference bundle post-trained with LoRA v2 on eight RoboTwin 2.0 tasks. The merged transformer/ can be used directly by wan_va_server.py --pretrained-model-path.

LoRA v2 uses rank 64, alpha 128, dropout 0.0, attention + FFN LoRA targets, and full training for action_embedder and action_proj_out.

Training tasks:

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

The bundle includes the base tokenizer/, text_encoder/, and vae/, plus the step 10000 merged transformer/ and lora_adapter/ for analysis.