# Portable model configuration. Set MR0_RELEASE_DIR to the downloaded repository root. run_id: Meituan-Robotics-0 run_root_dir: ./runs seed: 42 trackers: [] is_debug: false framework: name: QwenMMDiT qwenvl: base_vlm: ${oc.env:MR0_RELEASE_DIR,.}/base_vlm/Qwen3.5-4B-Action attn_implementation: flash_attention_2 vl_hidden_dim: 2560 action_model: action_model_type: MMDiT-Psi0 action_dim: 20 state_dim: 20 future_action_window_size: 15 action_horizon: 16 past_action_window_size: 0 repeated_diffusion_steps: 8 vl_feature_dim: 2560 num_target_vision_tokens: 32 add_pos_embed: true max_seq_len: 1024 noise_beta_alpha: 1.5 noise_beta_beta: 1.0 noise_s: 0.999 num_timestep_buckets: 1000 num_inference_timesteps: 10 state_placement: condition state_dropout_ratio: 0.0 prediction_type: sample t_eps: 0.05 datasets: vla_data: dataset_py: lerobot_datasets data_root_dir: ${oc.env:MR0_DATA_ROOT,./data} modality_root_dir: ${oc.env:MR0_DATA_ROOT,./data} data_mix: dualarm5_all_syn_no_a1_buy action_type: abs_eef action_mode: abs sequential_step_sampling: false lazy_step_sampling: true skip_empty_language_check: true balance_dataset_weights: true balance_trajectory_weights: false language_sampling: current_instruction: 0.5 task_description: 0.5 per_device_batch_size: 4 load_all_data_for_training: true max_pixels: 100352 min_pixels: 784 video_backend: torchcodec include_state: true num_workers: 4 trainer: epochs: 100 max_train_steps: 500000 num_warmup_steps: 5000 save_interval: 10000 eval_interval: 10000 learning_rate: base: 1.0e-05 qwen_vl_interface: 1.0e-05 action_model: 0.0001 lr_scheduler_type: cosine_with_min_lr scheduler_specific_kwargs: min_lr: 5.0e-07 freeze_modules: '' loss_scale: vla: 1.0 vlm: 0.1 warmup_ratio: 0.1 logging_frequency: 100 gradient_clipping: 1.0 gradient_accumulation_steps: 1 optimizer: name: AdamW betas: - 0.9 - 0.95 eps: 1.0e-08 weight_decay: 1.0e-08 is_resume: false resume_epoch: null resume_step: null pretrained_checkpoint: null enable_gradient_checkpointing: false enable_mixed_precision_training: true