| { |
| "global_step": 10000, |
| "effective_global_batch": 32, |
| "pretrained_model": "robotics-diffusion-transformer/rdt-1b", |
| "model_report": { |
| "trainable": 1207875712, |
| "total": 1244332160, |
| "percentage": 97.07019964829969, |
| "finetune_mode": "full", |
| "qwen_fusion": "hidden_cross_attention", |
| "plan_conditioning_trainable": true, |
| "rdt_state_dim": 128, |
| "state_adaptor_frozen": true, |
| "condition_adaptors_frozen": true, |
| "language_adaptor_frozen": true, |
| "image_adaptor_frozen": true, |
| "native_state_encoder": true, |
| "native_action_encoder": true, |
| "action_adaptor_trainable": false, |
| "lora_target_count": 0, |
| "pretrained": null, |
| "base_artifact": { |
| "format": "upstream_libero_rdt_ema", |
| "checkpoint": "output_3/checkpoints/RDT-1B-LIBERO-Base/ema/model.safetensors", |
| "loaded_tensors": 618, |
| "loaded_runner_interfaces": [ |
| "lang_adaptor", |
| "img_adaptor", |
| "state_adaptor" |
| ], |
| "reinitialized_tensors": [], |
| "adapted_tensors": [ |
| "model.lang_cond_pos_embed" |
| ], |
| "shape_mismatches": { |
| "model.lang_cond_pos_embed": [ |
| [ |
| 1, |
| 1024, |
| 2048 |
| ], |
| [ |
| 1, |
| 128, |
| 2048 |
| ] |
| ] |
| } |
| } |
| }, |
| "horizon_loss_weights": null, |
| "mask_noisy_gripper_input": true, |
| "training_objective": { |
| "horizon_loss_weights": null, |
| "mask_noisy_gripper_input": true, |
| "xyz_loss_weight": 0.0, |
| "gripper_bce_weight": 1.0, |
| "gripper_bce_logit_scale": 5.0, |
| "rotation_geodesic_weight": 0.0, |
| "qwen_fusion_loss_weight": 0.5, |
| "qwen_fusion_loss_margin": 0.002 |
| }, |
| "config": { |
| "seed": 42, |
| "rdt_repo": "/home/ubuntu/RoboticsDiffusionTransformer", |
| "pretrained_model": "robotics-diffusion-transformer/rdt-1b", |
| "output_dir": "/workspace/ThinkFlow-RDT-1B/output_2/libero_spatial_b0_hidden_from_libero_base_10k", |
| "model": { |
| "action_dim": 128, |
| "state_dim": 128, |
| "pred_horizon": 64, |
| "qwen_hidden_size": 2560, |
| "qwen_kv_dim": 2048, |
| "lang_token_dim": 4096, |
| "img_token_dim": 1152, |
| "max_lang_tokens": 128, |
| "image_tokens": 4374, |
| "hidden_size": 2048, |
| "depth": 28, |
| "num_heads": 32, |
| "dtype": "bfloat16", |
| "lang_adaptor": "mlp2x_gelu", |
| "img_adaptor": "mlp2x_gelu", |
| "state_adaptor": "mlp3x_gelu", |
| "copy_pretrained_final_fc1": true, |
| "gradient_checkpointing": false, |
| "finetune_mode": "full", |
| "qwen_fusion": "hidden_cross_attention", |
| "conditioning_variant": "b0", |
| "spatial_token_count": 1, |
| "waypoint_dim": 2, |
| "waypoint_embed_dim": 64, |
| "rdt_state_dim": 128, |
| "freeze_state_adaptor": true, |
| "freeze_condition_adaptors": true, |
| "freeze_language_adaptor": null, |
| "freeze_image_adaptor": null, |
| "state_encoder_layout": "rdt_native_128", |
| "action_encoder_layout": "rdt_native_128", |
| "native_rdt_128_mapping": "libero_joint_eef_delta", |
| "convert_cached_gripper_closed_to_open": false, |
| "allow_random_frozen_state_adaptor": false, |
| "pretrained_copy_mode": "compatible", |
| "cache_state_dim": 9, |
| "cache_action_dim": 7 |
| }, |
| "lora": { |
| "rank": 32, |
| "alpha": 64, |
| "dropout": 0.05, |
| "target_self_attention": true, |
| "target_cross_attention": true, |
| "target_ffn": true, |
| "train_final_layer": true |
| }, |
| "noise_scheduler": { |
| "num_train_timesteps": 1000, |
| "num_inference_timesteps": 5, |
| "beta_schedule": "squaredcos_cap_v2", |
| "prediction_type": "sample", |
| "clip_sample": false |
| }, |
| "data": { |
| "train_manifest": "/workspace/ThinkFlow-RDT-1B/output_2/libero_spatial_b0_hidden_from_libero_base_10k/manifests/train_manifest.jsonl", |
| "val_manifest": "/workspace/ThinkFlow-RDT-1B/output_2/libero_spatial_b0_hidden_from_libero_base_10k/manifests/val_manifest.jsonl", |
| "num_workers": 4, |
| "pin_memory": true, |
| "persistent_workers": true, |
| "excluded_dataset_ids": [], |
| "episode_aware_shuffle": true, |
| "shuffle_validation": false, |
| "stratified_validation": true, |
| "action_stats_paths": null |
| }, |
| "training": { |
| "max_steps": 10000, |
| "micro_batch_size": 8, |
| "gradient_accumulation_steps": 4, |
| "learning_rate_lora": 0.0001, |
| "learning_rate_interfaces": 0.0001, |
| "weight_decay_interfaces": 0.01, |
| "warmup_steps": 500, |
| "max_grad_norm": 1.0, |
| "log_every": 10, |
| "validate_every": 500, |
| "save_every": 1000, |
| "validation_batches": 8, |
| "sample_validation_batches": 1, |
| "mixed_precision": "bf16", |
| "report_to": "wandb", |
| "global_batch_size": 32, |
| "learning_rate": 0.0001, |
| "wandb_project": "ThinkLite B0 LIBERO", |
| "wandb_run_name": "libero-spatial-b0-hidden-10k", |
| "validation_seed": 12345, |
| "validation_samples": 256, |
| "skip_nonfinite_updates": true, |
| "max_consecutive_nonfinite_updates": 10, |
| "log_gradient_stats": true, |
| "qualitative_validation_examples": 32, |
| "validation_batch_size": 32, |
| "sampled_validation_horizons": [ |
| 10 |
| ], |
| "qwen_fusion_loss_weight": 0.5, |
| "qwen_fusion_loss_margin": 0.002, |
| "conditioning_warmup_steps": 0, |
| "conditioning_warmup_fusion_loss_weight": null, |
| "learning_rate_rdt_backbone": 0.0001, |
| "learning_rate_rdt_cross_attention": 0.0001, |
| "learning_rate_plan_projector": 0.0001, |
| "conditioning_warmup_cross_attention_learning_rate": 0.0001 |
| } |
| }, |
| "segment_stop_step": 10000, |
| "conditioning_interface": { |
| "type": "hidden_cross_attention", |
| "cached_kv_retained_but_unused": true, |
| "spatial_token_count": 1, |
| "qwen_hidden_size": 2560, |
| "rdt_condition_dim": 2048, |
| "qwen_token_selector": "think_end", |
| "conditioning_variant": "b0" |
| } |
| } |
