seed: 42
wandb_entity: null
wandb_project: Mem-0
is_debug: false
execution_module:
  qwen_vl:
    model_path: ./checkpoints/Qwen3-VL-2B-Instruct
    system_prompt: null
  memory_bank:
    window_size: 30
    initial_anchor_size: 1
    num_heads: 8
    dropout: 0.1
    memory_accumulation: 8
  action_model:
    action_model_type: DiT-B
    hidden_size: 2048
    add_pos_embed: true
    max_seq_len: 1024
    action_dim: 16
    state_dim: 16
    action_horizon: 30
    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: 8
    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: 2048
      positional_embeddings: null
  classifier:
    hidden_sizes:
    - 6144
    - 2048
    - 512
    dropout: 0.1
    pos_weight: 10.0
    focal_gamma: 1.0
    threshold: 0.5
  loss_weights:
    lambda_action: 1.0
    lambda_classifier: 0.2
dataloader:
  type: rmbench_random_episode
  num_workers: 2
  shuffle_episodes: true
  seed: 42
  infinite: true
  norm_stats_path: ./assets/m1_mix/norm_stats.json
trainer:
  wandb_run_name: mem0_exec_m1_mix
  resume: null
  batch_size: 56
  train_steps: 50000
  eval_action_interval: 100
  log_interval: 10
  classifier_metrics_window_size: 20
  enable_wandb: true
  checkpoint_dir: /mnt/data/limingleyang/rmbench_assets/checkpoints/m1_mix
  save_every_steps: 1000
  save_final: true
  scheduler: cosine
  warmup_ratio: 0.05
  grad_clip_norm: 2.5
  weight_decay: 0.005
  learning_rate:
    base: 1.0e-05
    qwen_model: 1.0e-05
    action_model: 0.0001
    classifier: 0.0001
  min_lr:
    base: 1.0e-06
    qwen_model: 1.0e-06
    action_model: 5.0e-06
    classifier: 5.0e-06
  freeze_modules: ''
  download_videos: false
vla_dataset:
  RMBench:
    repo_id: /mnt/data/limingleyang/rmbench_assets/lerobot_datasets/m1_mix
    features_to_load:
    - observation.state
    - action
    - subtask
    - subtask_end
    - episode_id
    - observation.image.head_camera