{
  "run_id": "0508_QwenLatent_13tasks_history_ActionState_50k",
  "run_root_dir": "./runs",
  "seed": 42,
  "trackers": [
    "jsonl",
    "wandb"
  ],
  "wandb_entity": "timsty",
  "wandb_project": "vla_jepa",
  "is_debug": false,
  "framework": {
    "name": "QwenLatent_history",
    "qwenvl": {
      "base_vlm": "/dataset_rc_mm/tangwl3@xiaopeng.com/jd/checkpoints/Qwen/Qwen3-VL-2B-Instruct",
      "attn_implementation": "flash_attention_2",
      "vl_hidden_dim": 2048,
      "num_data_tokens": 32
    },
    "action_model": {
      "ckpt_path": "/dataset_rc_mm/tangwl3@xiaopeng.com/jd/code/lap/runs/0418_Action_13tasks_actionstate_fixchunk15/final_model/pytorch_model.pt",
      "action_size": 37,
      "state_size": 74,
      "use_state": "${datasets.vla_data.state_use_action_chunk}",
      "hidden_size": 1024,
      "intermediate_size": 3072,
      "dataset_vocab_size": 256,
      "num_data_tokens": 32,
      "min_action_len": 5,
      "num_encoder_layers": 28,
      "num_decoder_layers": 28,
      "num_attention_heads": 16,
      "num_key_value_heads": 8,
      "head_dim": 128,
      "max_position_embeddings": 2048,
      "max_action_chunk_size": 50,
      "rms_norm_eps": 1e-06,
      "attention_dropout": 0.0,
      "use_vae_reparameterization": false,
      "use_ema": false,
      "chunk_size": "${datasets.vla_data.chunk_size}",
      "loss_mode": "full",
      "qwen3_pretrained_name_or_path": "/dataset_rc_mm/tangwl3@xiaopeng.com/jd/checkpoints/Qwen/Qwen3-0.6B"
    }
  },
  "datasets": {
    "vla_data": {
      "dataset_py": "lerobot_datasets",
      "data_root_dir": "/dataset_rc_mm/tangwl3@xiaopeng.com/jd/data/transfer/data/IPEC-COMMUNITY",
      "data_mix": "cross_embodiedment_13tasks",
      "CoT_prompt": "Task: {instruction}. What are the next 15 actions to take?",
      "default_image_resolution": [
        3,
        224,
        224
      ],
      "per_device_batch_size": 32,
      "load_all_data_for_training": true,
      "obs": [
        "image_0"
      ],
      "image_size": [
        224,
        224
      ],
      "video_backend": "torchcodec",
      "load_video": true,
      "chunk_size": 30,
      "state_use_action_chunk": true,
      "num_history_steps": 15,
      "include_state": "${datasets.vla_data.state_use_action_chunk}"
    }
  },
  "trainer": {
    "epochs": 100,
    "max_train_steps": 50000,
    "num_warmup_steps": 5000,
    "num_stable_steps": 0,
    "mode": "freeze_action_encoder_decay_aux_loss",
    "loss_weights_decay_steps": 5000,
    "save_interval": 5000,
    "eval_interval": 50,
    "max_checkpoints_to_keep": 10,
    "learning_rate": {
      "base": 5e-05,
      "qwen_vl_interface": 5e-05,
      "action_model": 5e-05
    },
    "lr_scheduler_type": "warmup_stable_cosine",
    "scheduler_specific_kwargs": {
      "min_lr_ratio": 0.001
    },
    "freeze_modules": "",
    "loss_scale": {
      "align_loss": 1.0,
      "recon_loss": 1.0,
      "predict_loss": 1.0
    },
    "warmup_ratio": 0.1,
    "weight_decay": 0.0,
    "logging_frequency": 10,
    "gradient_clipping": 5.0,
    "gradient_accumulation_steps": 1,
    "optimizer": {
      "name": "AdamW",
      "betas": [
        0.9,
        0.95
      ],
      "eps": 1e-08,
      "weight_decay": 1e-08
    },
    "is_resume": false,
    "resume_epoch": null,
    "resume_step": null,
    "enable_gradient_checkpointing": true,
    "enable_mixed_precision_training": true
  },
  "output_dir": "./runs/0508_QwenLatent_13tasks_history_ActionState_50k"
}