run_id: starvla_continual_libero_goal
run_root_dir: results/Checkpoints
seed: 42
trackers: [jsonl]
wandb_entity: ""
wandb_project: llavavla
is_debug: false


framework:
  name: QwenGR00T
  qwenvl:
    base_vlm: /share/lipengteng/VLA-Engine-Developer/data/pretrained_models/Qwen2.5-VL-3B-Instruct
    attn_implementation: flash_attention_2
    vl_hidden_dim: 2048
  dino:
    dino_backbone: dinov2_vits14
  action_model:
    action_model_type: DiT-B
    action_hidden_dim: 1024
    hidden_size: 1024
    add_pos_embed: True
    max_seq_len: 1024
    action_dim: 7
    state_dim: 7
    future_action_window_size: 7
    action_horizon: 8
    past_action_window_size: 0
    repeated_diffusion_steps: 8
    noise_beta_alpha: 1.5
    noise_beta_beta: 1.0
    noise_s: 0.999
    num_timestep_buckets: 1000
    num_inference_timesteps: 4
    num_target_vision_tokens: 32
    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

  reduce_in_full_precision: true


datasets:
  vla_data:
    dataset_py: lerobot_datasets
    data_root_dir: /share/lipengteng/VLA-Engine-Developer/data/datasets/LEROBOT_LIBERO_DATA
    dataset_mix: libero_goal    # base dataset suite for CL task splitting
    action_type: delta_qpos
    sequential_step_sampling: False
    CoT_prompt: "Your task is {instruction}. To identify the key objects for your task. Locate their bounding boxes in [x1,y1,x2,y2] format."
    CoT_answer: bbox
    default_image_resolution: [3, 224, 224]
    per_device_batch_size: 8
    load_all_data_for_training: true
    obs: ["image_0"]
    video_backend: torchvision_av


# ============================================================================
# Continual Learning Configuration
# ============================================================================
continual_learning:
  task_sequence: libero_goal      # CL sequence name (see continual_learning.py)
  steps_per_task: 10000            # training steps per task
  save_checkpoint_per_task: true  # save model after each task completes

  replay:
    enabled: true                 # set false to disable replay (naive sequential)
    method: experience_replay
    buffer_size_per_task: 1000     # samples to store per completed task (~20% of per-task data)
    replay_batch_ratio: 0.5       # fraction of batch from replay buffer
    balanced_sampling: true       # equal samples per task to prevent recency bias


# ============================================================================
# LoRA Configuration (optional, set enabled: false to use full fine-tuning)
# ============================================================================
lora:
  enabled: false                   # set true to use LoRA on VLM backbone
  rank: 32                         # LoRA rank
  alpha: 16                        # scaling factor
  dropout: 0.05
  target_modules: "all-linear"     # applied to VLM backbone linear layers


trainer:
  max_train_steps: 100000          # total budget (10 tasks x 10000)
  num_warmup_steps: 500
  save_interval: 5000
  eval_interval: 5000
  learning_rate:
    base: 2.5e-05
    qwen_vl_interface: 1.0e-05
    action_model: 1.0e-04
  lr_scheduler_type: cosine_with_min_lr
  scheduler_specific_kwargs:
    min_lr: 1.0e-06
  freeze_modules: ''
  loss_scale:
    vla: 1.0
  max_grad_norm: 1.0
  warmup_ratio: 0.1
  weight_decay: 0.0
  logging_frequency: 10
  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
  enable_gradient_checkpointing: true
  enable_mixed_precision_training: true