{
  "step": 20000,
  "args": {
    "base_model": "/data/base_model",
    "config_model": "/data/config_model",
    "dataset": "/data/dataset",
    "vl_model_path": null,
    "canonical_vl_model_name": "nvidia/Cosmos-Reason2-2B",
    "rl_label_dir": "offline_rl_labels_reward_v2_odom_grasp",
    "output_dir": "/data/output",
    "experiment_name": "task0025_subtask_lang_offline_rl_reward_v2_from_minmax60k_20k_gpu0_20260619_005429",
    "max_steps": 20000,
    "global_batch_size": 4,
    "gradient_accumulation_steps": 1,
    "num_workers": 4,
    "actor_lr": 2e-05,
    "critic_lr": 1e-05,
    "weight_decay": 1e-05,
    "warmup_ratio": 0.05,
    "beta": 0.05,
    "advantage_clip_max": 6.0,
    "advantage_clip_min": 0.25,
    "normalize_advantage": true,
    "critic_coef": 1.0,
    "critic_hidden_dim": 1024,
    "critic_warmup_steps": 1000,
    "warmup_actor_coef": 0.0,
    "sft_anchor_coef": 0.15,
    "arm_weight": 1.25,
    "wrist_weight": 1.75,
    "gripper_weight": 3.0,
    "odom_move_weight": 1.25,
    "odom_manip_weight": 2.5,
    "num_tasks": 3,
    "episode_sampling_rate": 0.1,
    "shard_size": 1024,
    "num_shards_per_epoch": 100000,
    "save_steps": 20000,
    "logging_steps": 10,
    "max_grad_norm": 1.0,
    "seed": 42,
    "bf16": true,
    "tf32": true,
    "local_files_only": false,
    "use_percentiles": false,
    "clip_outliers": true,
    "mask_splice_crossing": true,
    "save_optimizer_state": false,
    "save_final_model": true,
    "dry_run_data": false,
    "dry_run_forward": false
  }
}