action_dim: 7
action_model_type: DiT-B
data_root_dir: /scr/user/data
debug: false
future_action_window_size: 15
hf_token: hf_token
image_aug: true
is_resume: false
load_all_data_for_training: true
past_action_window_size: 3
pretrained_checkpoint: path/to/ckpt
repeated_diffusion_steps: 4
resume_epoch: null
resume_step: null
run_id: step3_HIS_Sil_prism-qwen25-dinosiglip-224px+bridge_rt_1+diffusion+n8+b16+x42--image_aug
run_id_note: null
run_root_dir: outputs/step3_HIS_ditb_8_lr2e-5_b16_fa0_shuffle_bridge_rt_1
save_interval: 2500
seed: 42
trackers:
  - jsonl
  - wandb
use_ema: false
vla:
  action_tokenizer: extra_action_tokenizer
  base_vlm: prism-qwen25-extra-dinosiglip-224px+0_5b
  data_mix: bridge_rt_1
  enable_gradient_checkpointing: true
  enable_mixed_precision_training: true
  epochs: 100
  expected_world_size: 32
  freeze_llm_backbone: false
  freeze_vision_backbone: false
  global_batch_size: 1024
  learning_rate: 4.0e-05
  lr_scheduler_type: constant
  max_grad_norm: 1.0
  max_steps: null
  per_device_batch_size: 32
  reduce_in_full_precision: true
  shuffle_buffer_size: 250000
  train_strategy: fsdp-full-shard
  type: prism-qwen25-dinosiglip-224px+0_5b
  unfreeze_last_llm_layer: false
  vla_id: prism-qwen25-dinosiglip-224px+0_5b
  warmup_ratio: 0.0
  weight_decay: 0.0
wandb_entity: ""
wandb_project: ""