| { |
| "step": 100000, |
| "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_v3_phase", |
| "output_dir": "/data/output", |
| "experiment_name": "task0047_phase_memory_offline_rl_from_chunk32_100k_80k_gpu0_20260702_014514", |
| "max_steps": 100000, |
| "global_batch_size": 8, |
| "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, |
| "use_memory_critic": true, |
| "memory_lags": "0,1,2,4,8,16,32,64", |
| "memory_state_dim": 16, |
| "memory_action_dim": 16, |
| "memory_token_dim": 256, |
| "memory_embedding_dim": 512, |
| "memory_num_layers": 1, |
| "memory_q_coef": 1.0, |
| "use_q_advantage": false, |
| "num_tasks": 5, |
| "num_phases": 25, |
| "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": true, |
| "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 |
| } |
| } |
