deformable-bench/lingbot-grasp-bag-strap-dr-v1

LingBot-VA grasp-bag-strap DR v1

LingBot-VA transformer fine-tuned on the deformable-bench/grasp_bag_strap_dr_v1 dataset.

Training setup

• Base model: robbyant/lingbot-va-base
• Dataset revision: 239700e0cc56136cdc03a1d3625c1064e6fd48b9
• Dataset size: 200 episodes and 83,566 frames
• Cameras: static, left hand, and right hand
• Action/state dimension: 14
• Latent extraction frame stride: 4
• Latent extraction shards: 8
• DataLoader workers: 16
• GPUs: 8 H20 GPUs
• Batch size: 1 per rank
• Learning rate: 1e-5
• Precision: bf16
• CFG dropout probability: 0.1
• Training length: 30,000 optimizer steps

Training loss

The run completed normally at step 30,000. Mean logged losses over selected windows were:

Window	Latent	Action
Steps 10-100	0.202769	0.024912
Steps 14,901-15,000	0.025104	0.000480
Final 200 steps	0.010343	0.000362
Final logged step	0.008182	0.000339

Both latent and action losses converged steadily without a late rebound.

Files

• transformer/diffusion_pytorch_model.safetensors: final transformer at step 30,000
• transformer/config.json: transformer architecture configuration
• train.log: complete training log

Only the fine-tuned transformer is included. The tokenizer, text encoder, and VAE should be loaded from robbyant/lingbot-va-base.

Intended use

This checkpoint is intended for research and evaluation with the matching LingBot-VA codebase and the grasp_bag_strap_dr_v1 observation/action schema. It has not been validated for safety-critical or out-of-distribution deployment.