datasets:
  vla_data:
    data_mix: libero_all
    data_root_dir: /home/jye624/Datasets/LIBERO
    dataset_py: lerobot_datasets
    per_device_batch_size: 8
    sequential_step_sampling: false
    video_backend: torchvision_av
framework:
  name: CosmoPredict2GR00T
  action_model:
    action_dim: 7
    action_horizon: 8
    action_model_type: DiT-B
    add_pos_embed: true
    diffusion_model_cfg:
      cross_attention_dim: 2048
      dropout: 0.2
      final_dropout: true
      interleave_self_attention: true
      norm_type: ada_norm
      num_layers: 16
      output_dim: 1024
      positional_embeddings: null
    future_action_window_size: 7
    hidden_size: 1024
    max_seq_len: 1024
    noise_beta_alpha: 1.5
    noise_beta_beta: 1.0
    noise_s: 0.999
    num_inference_timesteps: 4
    num_target_vision_tokens: 32
    num_timestep_buckets: 1000
    past_action_window_size: 0
    repeated_diffusion_steps: 8
    state_dim: 7
  obs_image_size: null
  qwenvl:
    base_vlm: /home/jye624/Models/Pretrained_models/Qwen3-VL-4B-Instruct
  world_model:
    base_wm: ./playground/Pretrained_models/nvidia/Cosmos-Predict2-2B-Video2World
    extract_layers:
    - -1
output_dir: ./results/Checkpoints/0405_libero4in1_CosmoPredict2GR00T
run_id: 0405_libero4in1_CosmoPredict2GR00T
run_root_dir: ./results/Checkpoints
seed: 42
trainer:
  eval_interval: 100
  freeze_modules: true
  gradient_accumulation_steps: 1
  gradient_clipping: 1.0
  is_resume: false
  learning_rate:
    action_model: 0.0001
    base: 2.5e-05
    qwen_vl_interface: 1.0e-05
  logging_frequency: 100
  lr_scheduler_type: cosine_with_min_lr
  max_train_steps: 80000
  num_warmup_steps: 5000
  optimizer:
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-08
  save_interval: 10000
  scheduler_specific_kwargs:
    min_lr: 1.0e-06
wandb_entity: jinhuiye
wandb_project: starVLA_Libero