# pot14 自采数据微调配置（单活动臂 7 维 + 三路相机）
# 基于 robotwin 骨架，数据来自 scripts/convert_pot14_to_lerobot.py

model:
  model_path: ../models/pretrained/lingbot-vla-4b
  tokenizer_path: ../models/pretrained/Qwen2.5-VL-3B-Instruct

data:
  datasets_type: vla
  data_name: pot14
  train_path: ../data/processed/pot14_right_arm
  joints:
    - arm.position: 6
    - effector.position: 1
  cameras:
    - camera_top
    - camera_wrist_left
    - camera_wrist_right
  num_workers: 4
  pin_memory: true
  prefetch_factor: 2
  video_backend: torchcodec
  norm_type: meanstd
  norm_stats_file: assets/norm_stats/pot14_right_arm.json

train:
  output_dir: "output/pot14/"
  data_parallel_mode: fsdp2
  enable_full_shard: false
  module_fsdp_enable: true
  use_compile: true
  rmpad: false
  rmpad_with_pos_ids: false
  ulysses_parallel_size: 1
  freeze_vision_encoder: false
  tokenizer_max_length: 72
  max_action_dim: 75
  max_state_dim: 75
  lr: 5.0e-5
  lr_decay_style: constant
  micro_batch_size: 2
  gradient_accumulation_steps: 16
  max_steps: 1024
  ckpt_manager: dcp
  save_steps: 0
  save_epochs: 0
  max_checkpoints_to_keep: 1
  save_optimizer: false
  enable_fp32: true
  enable_resume: false