| { |
| "feature_type": "latent_student_spatial_kv", |
| "config": "configs/part3_rdt1b.yaml", |
| "root": "/workspace/ThinkFlow-RDT-1B/dataset/hf_parts/part_1", |
| "splits": [ |
| "validation", |
| "train", |
| "test" |
| ], |
| "datasets": [ |
| "bc_z", |
| "bridge", |
| "droid", |
| "fractal", |
| "kuka" |
| ], |
| "dataset_schedule": "round_robin", |
| "seed": 42, |
| "stage": null, |
| "stage_count": 3, |
| "droid_stage_count": 2, |
| "split_ratios": [ |
| 0.8, |
| 0.1, |
| 0.1 |
| ], |
| "max_samples_per_split": null, |
| "max_train_samples": null, |
| "max_validation_samples": null, |
| "max_test_samples": null, |
| "normalize_actions": true, |
| "action_target_mode": "delta", |
| "pred_horizon": 64, |
| "state_dim": 7, |
| "action_dim": 7, |
| "cache_state_dim": 7, |
| "cache_action_dim": 7, |
| "rdt_state_dim": 128, |
| "rdt_action_dim": 128, |
| "state_encoder_layout": "rdt_native_128", |
| "action_encoder_layout": "rdt_native_128", |
| "max_samples_per_episode": 32, |
| "require_exact_samples_per_episode": false, |
| "gripper_change_scope": "first_directional", |
| "open_to_close_before": 4, |
| "open_to_close_after": 4, |
| "close_to_open_before": 4, |
| "close_to_open_after": 4, |
| "student_model_id": "/workspace/model/LatentStudent-ckpt-400-fixed", |
| "processor_id": "/workspace/model/stage1_unsloth", |
| "spatial_parameters_path": null, |
| "latent_student_code_dir": "/workspace/VLA-FYP/train/stage2", |
| "latent_count": 6, |
| "spatial_token_count": 5, |
| "layer_index": 7, |
| "latent_student_contract": { |
| "hidden_size": 2560, |
| "num_hidden_layers": 32, |
| "num_key_value_heads": 4, |
| "head_dim": 256, |
| "layer_index": 7, |
| "layer_type": "full_attention", |
| "flattened_kv_dim": 2048, |
| "latent_count": 6, |
| "spatial_token_count": 5, |
| "end_think_token_id": 248069 |
| }, |
| "prompt_template": "You are a robot manipulation assistant. Given an observation image and a task instruction, predict the end-effector's 2D trajectory as 5 waypoints. Output ONLY the coordinate list in this exact format: [[x1,y1],[x2,y2],[x3,y3],[x4,y4],[x5,y5]]\n\nTask: The task is {task}. What is the trajectory that the end effector should take?", |
| "qwen_kv_dim": 2048, |
| "qwen_batch_size": 32, |
| "qwen_cache_scope": "per_sample", |
| "qwen_kv_granularity": "five_spatial_tokens_per_sample_step", |
| "include_t5": true, |
| "t5_model_id": "google/t5-v1_1-xxl", |
| "t5_precision": "bf16", |
| "save_padded_features": false, |
| "image_history_size": 2, |
| "max_images_per_sample": 6, |
| "image_storage_codec": "jpeg", |
| "image_storage_lossless": false, |
| "image_jpeg_quality": 90, |
| "cache_layout": "episode_pack", |
| "episode_shards_per_directory": 500, |
| "manifest_write_mode": "durable_per_episode" |
| } |
|
|
