| action_dim: 7 | |
| action_model_type: DiT-B | |
| data_root_dir: s3://real_data_raw/open_x_embodiment_origin | |
| debug: false | |
| disable_instruction: false | |
| fix_system1: false | |
| future_action_window_size: 15 | |
| hf_token: hf_token | |
| image_aug: true | |
| is_resume: false | |
| load_all_data_for_training: true | |
| num_of_meta_query: 64 | |
| past_action_window_size: 0 | |
| pretrained_checkpoint: null | |
| repeated_diffusion_steps: 4 | |
| resume_epoch: null | |
| resume_step: null | |
| run_id: sys12_meta_query_action_only_sync_pretraining_v2_query_64_mlp_lora_reimplement_history=1_state--image_augstage1 | |
| run_id_note: null | |
| run_root_dir: outputs/code_reimp | |
| save_interval: 1500 | |
| seed: 42 | |
| stage: stage1 | |
| trackers: | |
| - jsonl | |
| - wandb | |
| use_ema: false | |
| use_mm: false | |
| vla: | |
| action_tokenizer: extra_action_tokenizer | |
| base_vlm: /mnt/petrelfs/yangshuai1/yangshuai1/share_mllm/Eagle2-2B | |
| data_mix: bridge_rt_1 | |
| enable_gradient_checkpointing: true | |
| enable_mixed_precision_training: true | |
| epochs: 100 | |
| expected_world_size: 64 | |
| freeze_llm_backbone: false | |
| freeze_vision_backbone: false | |
| global_batch_size: 896 | |
| learning_rate: 5.0e-05 | |
| lr_scheduler_type: constant | |
| max_grad_norm: 1.0 | |
| max_steps: null | |
| per_device_batch_size: 14 | |
| reduce_in_full_precision: true | |
| shuffle_buffer_size: 250000 | |
| train_strategy: fsdp-shard-grad-op | |
| type: prism-qwen25-dinosiglip-224px+0_5b | |
| unfreeze_last_llm_layer: false | |
| vla_id: prism-qwen25-dinosiglip-224px+0_5b | |
| warmup_ratio: 0.0 | |
| weight_decay: 0.0 | |
| wandb_entity: shuaiyang2003 | |
| wandb_project: dual_sys | |