thingvellir's picture
Use complete VLA checkpoint; remove redundant Qwen weights
1337089 verified
Raw History Blame Contribute Delete
967 Bytes
{
"display_name": "Meituan-Robotics-0-RoboTwin",
"framework": "QwenMMDiT",
"checkpoint": "checkpoints/model.pt",
"original_checkpoint_basename": "steps_70000_pytorch_model.pt",
"training_step": 70000,
"action_dim": 14,
"action_horizon": 100,
"code_repository": "https://github.com/meituan/Meituan-Robotics-0",
"code_commit": "88c667933dfc941a71c9b142eb84b716ac4f0707",
"serialization": "PyTorch state_dict; unchanged from the selected source checkpoint",
"evaluation": {
"demo_clean_mean_success_percent": 93.22,
"demo_randomized_mean_success_percent": 93.88,
"tasks_per_setting": 50,
"episodes_per_task": 100,
"instruction_split": "unseen"
},
"repo_id": "MeituanRobotics/Meituan-Robotics-0-RoboTwin",
"checkpoint_contains": [
"Qwen3.5-4B backbone",
"MMDiT action head"
],
"backbone_metadata": "base_vlm/Qwen3.5-4B-Action/",
"separate_qwen_weight_file_required": false,
"checkpoint_tensor_count": 929
}