# INSIGHT GR00T N1.7 training config
# INSIGHT wrapper-only fields are stripped before LeRobot config parsing.
insight:
  checkpoint_interval: 5K
  checkpoint_slots: 2
  # Options: joint_abs | ee_abs_quat | ee_abs_rot6d | ee_delta
  action_space: ee_abs_rot6d
  # Options: coarse | detailed
  prompt_set: detailed
  save_steps: ["10K", "30K", "60K"]

dataset:
  repo_id: Whalswp/INSIGHTfixposV4_filtered_multispace_v2
  root: ${HOME}/INSIGHTfixposV4_filtered_multispace_v2

policy:
  type: groot_rkd
  device: cuda
  chunk_size: 16
  n_action_steps: 16
  push_to_hub: false
  # Trainable scope. This baseline trains projector, diffusion, and VLLN; LLM/vision stay frozen.
  # VLM
  tune_llm: false
  tune_visual: false
  tune_top_llm_layers: 0
  # Action Expert
  action_expert_num_layers: 8
  tune_projector: true
  tune_diffusion_model: true
  tune_vlln: true
  # RKD policy is used only to expose the shared DiT-depth setting.
  # This remains the action-loss-only control.
  rkd_enabled: false


seed: 42
batch_size: 64
steps: 60000
log_freq: 200
output_dir: ${HOME}/groot_insight/Abs_6D/DiT_Layer/8/Baseline
job_name: INSIGHT_6D_DiT8_Baseline

wandb:
  enable: true
  disable_artifact: true

# Omitted because these match LeRobot/GR00T defaults or are unused in this setup:
#
# dataset:
#   eval_split: 0.0
#   image_transforms:
#     enable: false  # default; set true only for train-image augmentation.
#
# policy:
#   base_model_path: nvidia/GR00T-N1.7-3B
#   embodiment_tag: new_embodiment
#   use_relative_actions: false
#   use_bf16: true
#   repo_id: Whalswp/INSIGHT  # unused while push_to_hub is false
#
# save_checkpoint: true
# use_policy_training_preset: true
# eval_steps: 0
# env_eval_freq: 0
#
# insight.prompt_set options:
# coarse: High-level task text that requires following the visual guide.
# detailed: Direction- and interaction-specific text, including rotation and push/pull details.