{ "title": "Ropedia Xperience-10M Current Result Versions and Model Groups", "generated_at_utc": "2026-06-11T04:42:46+00:00", "status": "pass", "version_count": 3, "model_group_count": 5, "comparison_rule": "Compare only rows with the same scope and target. Single-episode raw-feature metrics, 128-episode metadata baselines, Qwen3 structured JSON metrics, and the two Cosmos3 targets answer different questions: Nano future-window retrieval versus Super structured JSON Reasoner evaluation.", "version_reading_notes": [ "Version 1 is the public-sample 12-task harness with minimal and neural heads.", "Version 2 is the selected 128-episode same-split simple/NN baseline alignment.", "Version 3 is the verified model-branch layer: the current final Qwen3-Omni LoRA package is the JSON-task diagnostic result, Cosmos3-Nano is a future-window compatibility result, Cosmos3-Super Reasoner is a base-weight JSON-task evaluation, and Cosmos3-Super Forward-Dynamics LoRA is the first Super fine-tuned adapter branch." ], "versions": [ { "id": "v1_single_episode_public_sample", "title": "Single-Episode Public-Sample Task Suite", "status": "verified", "scope": "one public Xperience-10M sample episode", "source": "results/episode_task_suite/summary_report.json", "split": "chronological 70/30 within one episode", "counts": { "episodes": 1, "windows": 1161, "frames": 5821, "feature_dim": 8546, "task_count": 12, "neural_task_count": 12 }, "models": [ "minimal task heads", "compact neural MLP task heads" ], "task_metrics": [ { "task": "caption_grounding", "task_display_name": "Language Grounding", "simple_status": "pass", "simple_primary_metric": "mrr", "simple_primary_score": 0.016023479050338015, "neural_status": "pass", "neural_primary_metric": "mrr", "neural_primary_score": 0.01684125567132316 }, { "task": "contact_prediction", "task_display_name": "Contact State Prediction", "simple_status": "pass", "simple_primary_metric": "macro_f1", "simple_primary_score": 1.0, "neural_status": "pass", "neural_primary_metric": "macro_f1", "neural_primary_score": 1.0 }, { "task": "cross_modal_retrieval", "task_display_name": "Cross-Modal Retrieval", "simple_status": "pass", "simple_primary_metric": "mrr", "simple_primary_score": 0.26925966892956127, "neural_status": "pass", "neural_primary_metric": "mrr", "neural_primary_score": 0.1299971898648288 }, { "task": "hand_trajectory_forecast", "task_display_name": "Hand Trajectory Forecasting", "simple_status": "pass", "simple_primary_metric": "mpjpe", "simple_primary_score": 0.8646570444107056, "neural_status": "pass", "neural_primary_metric": "mpjpe", "neural_primary_score": 0.10785018652677536 }, { "task": "misalignment_detection", "task_display_name": "Multimodal Synchronization Detection", "simple_status": "pass", "simple_primary_metric": "f1", "simple_primary_score": 0.5051698670605613, "neural_status": "pass", "neural_primary_metric": "f1", "neural_primary_score": 0.7152682255845944 }, { "task": "modality_reconstruction", "task_display_name": "Cross-Modal Reconstruction", "simple_status": "pass", "simple_primary_metric": "r2", "simple_primary_score": -0.015271898913936655, "neural_status": "pass", "neural_primary_metric": "r2", "neural_primary_score": -0.010171410134180991 }, { "task": "next_action", "task_display_name": "Next-Action Prediction", "simple_status": "pass", "simple_primary_metric": "macro_f1", "simple_primary_score": 0.05925925925925927, "neural_status": "pass", "neural_primary_metric": "macro_f1", "neural_primary_score": 0.04186046511627907 }, { "task": "object_relevance", "task_display_name": "Object Relevance Prediction", "simple_status": "pass", "simple_primary_metric": "micro_f1", "simple_primary_score": 0.18034382095361662, "neural_status": "pass", "neural_primary_metric": "micro_f1", "neural_primary_score": 0.1679279279279279 }, { "task": "temporal_order", "task_display_name": "Temporal Order Verification", "simple_status": "pass", "simple_primary_metric": "accuracy", "simple_primary_score": 0.4540229885057471, "neural_status": "pass", "neural_primary_metric": "accuracy", "neural_primary_score": 0.8577586206896551 }, { "task": "timeline_action", "task_display_name": "Action Recognition", "simple_status": "pass", "simple_primary_metric": "macro_f1", "simple_primary_score": 0.05, "neural_status": "pass", "neural_primary_metric": "macro_f1", "neural_primary_score": 0.014814814814814814 }, { "task": "timeline_subtask", "task_display_name": "Procedure Step Recognition", "simple_status": "pass", "simple_primary_metric": "macro_f1", "simple_primary_score": 0.05056355513846935, "neural_status": "pass", "neural_primary_metric": "macro_f1", "neural_primary_score": 0.02810810810810811 }, { "task": "transition_detection", "task_display_name": "Action Boundary Detection", "simple_status": "pass", "simple_primary_metric": "macro_f1", "simple_primary_score": 0.6118237590630229, "neural_status": "pass", "neural_primary_metric": "macro_f1", "neural_primary_score": 0.5862068965517241 } ], "interpretation": "This layer verifies the 12 task contracts and raw multimodal feature pipeline on the public sample. It is not a cross-episode benchmark." }, { "id": "v2_multi_episode_128_aligned_metadata_baselines", "title": "128-Episode Aligned Simple/NN Baselines", "status": "pass", "scope": "selected 128-episode 96/16/16 split", "source": "results/omni_finetune/multi_episode_128_task_baselines/BASELINE_ALIGNMENT_REPORT.md", "split": "train/val/test by selected episode/session", "counts": { "rows": 3808, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "episode_counts": { "test": 16, "train": 96, "val": 16 }, "task_count": 12, "simple_supported_task_count": 8, "neural_supported_task_count": 6 }, "models": [ "metadata/text simple baselines", "metadata/text neural MLP baselines" ], "task_metrics": [ { "task": "timeline_action", "task_display_name": "Action Recognition", "simple_status": "pass", "simple_primary_metric": "macro_f1", "simple_primary_score": 0.00017511601435951318, "neural_status": "pass", "neural_primary_metric": "macro_f1", "neural_primary_score": 0.0 }, { "task": "timeline_subtask", "task_display_name": "Procedure Step Recognition", "simple_status": "pass", "simple_primary_metric": "macro_f1", "simple_primary_score": 0.0, "neural_status": "pass", "neural_primary_metric": "macro_f1", "neural_primary_score": 0.0 }, { "task": "transition_detection", "task_display_name": "Action Boundary Detection", "simple_status": "pass", "simple_primary_metric": "macro_f1", "simple_primary_score": 0.5219803670507895, "neural_status": "pass", "neural_primary_metric": "macro_f1", "neural_primary_score": 0.45822172492907925 }, { "task": "next_action", "task_display_name": "Next-Action Prediction", "simple_status": "pass", "simple_primary_metric": "macro_f1", "simple_primary_score": 0.00019966057701906761, "neural_status": "pass", "neural_primary_metric": "macro_f1", "neural_primary_score": 0.0 }, { "task": "hand_trajectory_forecast", "task_display_name": "Hand Trajectory Forecasting", "simple_status": "unsupported_without_raw_128_feature_blocks", "simple_primary_metric": "mpjpe", "simple_primary_score": null, "neural_status": "not_run", "neural_primary_metric": "", "neural_primary_score": null }, { "task": "contact_prediction", "task_display_name": "Contact State Prediction", "simple_status": "pass", "simple_primary_metric": "macro_f1", "simple_primary_score": 0.5167950693374422, "neural_status": "pass", "neural_primary_metric": "macro_f1", "neural_primary_score": 0.21951219512195122 }, { "task": "object_relevance", "task_display_name": "Object Relevance Prediction", "simple_status": "pass", "simple_primary_metric": "micro_f1", "simple_primary_score": 0.18221614227086183, "neural_status": "pass", "neural_primary_metric": "micro_f1", "neural_primary_score": 0.1053878034339846 }, { "task": "caption_grounding", "task_display_name": "Language Grounding", "simple_status": "pass", "simple_primary_metric": "mrr", "simple_primary_score": 0.012785504572093487, "neural_status": "not_run", "neural_primary_metric": "", "neural_primary_score": null }, { "task": "cross_modal_retrieval", "task_display_name": "Cross-Modal Retrieval", "simple_status": "unsupported_without_raw_128_feature_blocks", "simple_primary_metric": "mrr", "simple_primary_score": null, "neural_status": "not_run", "neural_primary_metric": "", "neural_primary_score": null }, { "task": "modality_reconstruction", "task_display_name": "Cross-Modal Reconstruction", "simple_status": "unsupported_without_raw_128_feature_blocks", "simple_primary_metric": "r2", "simple_primary_score": null, "neural_status": "not_run", "neural_primary_metric": "", "neural_primary_score": null }, { "task": "temporal_order", "task_display_name": "Temporal Order Verification", "simple_status": "pass", "simple_primary_metric": "f1", "simple_primary_score": 0.32713178294573647, "neural_status": "not_run", "neural_primary_metric": "", "neural_primary_score": null }, { "task": "misalignment_detection", "task_display_name": "Multimodal Synchronization Detection", "simple_status": "unsupported_without_raw_128_feature_blocks", "simple_primary_metric": "f1", "simple_primary_score": null, "neural_status": "not_run", "neural_primary_metric": "", "neural_primary_score": null } ], "interpretation": "This layer aligns the previous simple and neural baseline framing to the same selected 96/16/16 split used by the model branches. It uses public-safe JSONL metadata/text features, so raw-feature-only tasks remain explicitly unsupported until 128-run sensor feature blocks exist." }, { "id": "v3_multi_episode_foundation_model_branches", "title": "128-Episode Foundation-Model Branches", "status": "partial_verified", "scope": "selected 128-episode split and compatible derived windows", "source": "results/omni_finetune/verified_public/", "split": "episode/session held-out split; exact task target depends on backbone contract", "counts": { "verified_branch_count": 9, "qwen3_verified_package_count": 6, "cosmos3_verified_package_count": 3, "cosmos3_nano_verified_package_count": 1, "cosmos3_super_verified_package_count": 2 }, "models": [ "Qwen3-Omni LoRA", "Cosmos3-Nano future-window compatibility branch", "Cosmos3-Super Reasoner base-weight evaluation", "Cosmos3-Super forward-dynamics LoRA" ], "branches": [ { "id": "xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full", "title": "Cosmos3-Nano Future-Window World Model", "status": "verified", "backbone": "cosmos_world_model", "dataset_contract": "xperience10m_future_window_world_model_v0", "training_objective": "future_window_and_action_conditioned_world_modeling", "source": "results/omni_finetune/verified_public/xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full/verified_result_summary.json", "dataset_run_id": "xperience10m_cosmos3_nano_128ep_future_window_h5_compat", "train_run_id": "xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter", "eval_run_id": "xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full", "counts": { "dataset_samples": 3213, "dataset_episodes": 119, "split_counts": { "train": 2403, "test": 378, "val": 432 }, "train_samples": 2403, "val_samples": 432, "eval_samples": 378, "held_out_episode_count": 14, "num_processes": 1 }, "primary_metrics": { "future_retrieval_mrr": 0.022138720585222767, "future_retrieval_recall_at_5": 0.015873015873015872, "temporal_consistency": 0.09523809523809523, "feature_reconstruction_error": 3479.218317102503, "transition_accuracy": 0.9682539682539683, "contact_accuracy": 0.7433862433862434, "held_out_episode_count": 14 }, "history": [ { "epoch": 0, "train_loss": null, "val_loss": null, "note": "closed-form mean-delta adapter; no Cosmos diffusion weights fine-tuned in this compatibility run" } ], "is_current": true, "weights_repository": "planned separate Cosmos3 model repo after a real Cosmos diffusion/LoRA fine-tune exists; current result remains artifacts-only" }, { "id": "xperience10m_cosmos3_super_forward_dynamics_lora_128ep_train1epoch_256_attn_full8gpu_20260608_eval_test_full_fsdp", "title": "Cosmos3-Super Forward-Dynamics LoRA", "status": "verified", "backbone": "cosmos3_super_forward_dynamics", "dataset_contract": "xperience10m_camera_pose_forward_dynamics_v1", "training_objective": "camera_pose_conditioned_future_vision_velocity_lora", "source": "results/omni_finetune/verified_public/xperience10m_cosmos3_super_forward_dynamics_lora_128ep_train1epoch_256_attn_full8gpu_20260608_eval_test_full_fsdp/verified_result_summary.json", "dataset_run_id": "xperience10m_cosmos3_camera_pose_targets_20260608", "train_run_id": "xperience10m_cosmos3_super_forward_dynamics_lora_128ep_train1epoch_256_attn_full8gpu_20260608", "eval_run_id": "xperience10m_cosmos3_super_forward_dynamics_lora_128ep_train1epoch_256_attn_full8gpu_20260608_eval_test_full_fsdp", "counts": { "dataset_samples": 3808, "dataset_episodes": 119, "split_counts": { "test": 448, "train": 2848, "val": 512 }, "train_samples": 2848, "val_samples": 512, "eval_samples": 448, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "adapter_parameter_numel": 26214400, "held_out_episode_count": 14, "test_forward_dynamics_mse": 3.6853174321087345, "train_final_loss": 1.0785235166549683, "val_forward_dynamics_mse": 4.008244896889664 }, "history": [ { "epoch": 1, "note": "FSDP 8-GPU LoRA over camera-pose-conditioned future vision velocity loss; adapter weights are excluded from this public package.", "train_loss": 1.0785235166549683, "val_loss": 4.008244896889664 } ], "is_current": true, "weights_repository": "https://huggingface.co/cy0307/ropedia-cosmos3-super-forward-dynamics-lora-128ep" }, { "id": "xperience10m_cosmos3_super_reasoner_128ep_test_full_20260607", "title": "Cosmos3-Super Reasoner", "status": "verified", "backbone": "cosmos3_super_reasoner", "dataset_contract": "xperience10m_episode_json_qa_v1", "training_objective": "zero_shot_structured_episode_understanding_json_qa_via_vllm_reasoner", "source": "results/omni_finetune/verified_public/xperience10m_cosmos3_super_reasoner_128ep_test_full_20260607/verified_result_summary.json", "dataset_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605", "train_run_id": "xperience10m_cosmos3_super_reasoner_base_vllm_8gpu_20260607", "eval_run_id": "xperience10m_cosmos3_super_reasoner_128ep_test_full_20260607", "counts": { "dataset_samples": 3808, "dataset_episodes": 119, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "train_samples": 2848, "val_samples": 512, "eval_samples": 448, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "json_validity_rate": 0.5111607142857143, "action_macro_f1": 0.0008284021201089245, "subtask_accuracy": 0.0, "transition_accuracy": 0.36830357142857145, "next_action_accuracy": 0.013392857142857142, "contact_accuracy": 0.32142857142857145, "object_micro_f1": 0.13704276146316333, "held_out_episode_count": 14 }, "history": [], "is_current": true, "weights_repository": "none for this run: staged base nv-community/Cosmos3-Super weights were evaluated through vLLM; create a separate repo only after new adapter or fine-tuned weights exist" }, { "id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_eval", "title": "Qwen3-Omni LoRA", "status": "verified", "backbone": "qwen3_omni_lora", "dataset_contract": "xperience10m_episode_json_qa_v1", "training_objective": "structured_episode_understanding_json_qa", "source": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_eval/verified_result_summary.json", "dataset_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605", "train_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_lora", "eval_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_eval", "counts": { "dataset_samples": 3808, "dataset_episodes": 119, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "train_samples": 2848, "val_samples": 512, "eval_samples": 448, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "json_validity_rate": 0.875, "action_macro_f1": 0.0026621494447581404, "subtask_accuracy": 0.006696428571428571, "transition_accuracy": 0.8504464285714286, "next_action_accuracy": 0.024553571428571428, "contact_accuracy": 0.6450892857142857, "object_micro_f1": 0.22299431459254582, "held_out_episode_count": 14 }, "history": [ { "epoch": 1, "train_loss": 0.41304643672440994, "val_loss": 0.0330660454928875, "global_step": 356 } ], "is_current": false, "weights_repository": "historical diagnostic package; keep separate from the final 128-episode adapter repo" }, { "id": "xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu_lora_fsdp_full_train_noval_tail_logits_fullstatesave_v6_eval_test_full", "title": "Qwen3-Omni LoRA", "status": "verified", "backbone": "qwen3_omni_lora", "dataset_contract": "xperience10m_episode_json_qa_v1", "training_objective": "structured_episode_understanding_json_qa", "source": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu_lora_fsdp_full_train_noval_tail_logits_fullstatesave_v6_eval_test_full/verified_result_summary.json", "dataset_run_id": "xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu", "train_run_id": "xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu_lora_fsdp_full_train_noval_tail_logits_fullstatesave_v6", "eval_run_id": "xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu_lora_fsdp_full_train_noval_tail_logits_fullstatesave_v6_eval_test_full", "counts": { "dataset_samples": 3808, "dataset_episodes": 119, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "train_samples": 2848, "val_samples": 0, "eval_samples": 448, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "json_validity_rate": 0.8526785714285714, "action_macro_f1": 0.00213753459655099, "subtask_accuracy": 0.004464285714285714, "transition_accuracy": 0.828125, "next_action_accuracy": 0.022321428571428572, "contact_accuracy": 0.6517857142857143, "object_micro_f1": 0.23062730627306272, "held_out_episode_count": 14 }, "history": [ { "epoch": 1, "train_loss": 0.4121775626560694, "val_loss": null, "global_step": 356 } ], "is_current": false, "weights_repository": "historical diagnostic package; keep separate from the final 128-episode adapter repo" }, { "id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora_eval_test_full", "title": "Qwen3-Omni LoRA", "status": "verified", "backbone": "qwen3_omni_lora", "dataset_contract": "xperience10m_episode_json_qa_v1", "training_objective": "structured_episode_understanding_json_qa", "source": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora_eval_test_full/verified_result_summary.json", "dataset_run_id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora", "train_run_id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora", "eval_run_id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora_eval_test_full", "counts": { "dataset_samples": 34269, "dataset_episodes": 119, "split_counts": { "test": 4032, "train": 25629, "val": 4608 }, "train_samples": 25629, "val_samples": 1024, "eval_samples": 4032, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "json_validity_rate": 1.0, "action_macro_f1": 0.002289711036077459, "subtask_accuracy": 0.011194029850746268, "transition_accuracy": 0.9908234126984127, "next_action_accuracy": 0.053618594823032224, "contact_accuracy": 0.7864583333333334, "object_micro_f1": 0.31614599936244814, "held_out_episode_count": 14 }, "history": [ { "epoch": 1, "train_loss": 0.06255286606544624, "val_loss": 0.02668904885649681, "global_step": 3204 } ], "is_current": true, "weights_repository": "https://huggingface.co/cy0307/ropedia-qwen3-omni-lora-128ep" }, { "id": "xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora_eval_test_full", "title": "Qwen3-Omni LoRA", "status": "verified", "backbone": "qwen3_omni_lora", "dataset_contract": "xperience10m_episode_json_qa_v1", "training_objective": "structured_episode_understanding_json_qa", "source": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora_eval_test_full/verified_result_summary.json", "dataset_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605", "train_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora", "eval_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora_eval_test_full", "counts": { "dataset_samples": 3808, "dataset_episodes": 119, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "train_samples": 2848, "val_samples": 512, "eval_samples": 448, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "json_validity_rate": 0.9977678571428571, "action_macro_f1": 0.0024331644885523347, "subtask_accuracy": 0.002232142857142857, "transition_accuracy": 0.9709821428571429, "next_action_accuracy": 0.029017857142857144, "contact_accuracy": 0.71875, "object_micro_f1": 0.30160427807486634, "held_out_episode_count": 14 }, "history": [ { "epoch": 1, "train_loss": 0.41282760031950355, "val_loss": 0.03288277983665466, "global_step": 356 }, { "epoch": 2, "train_loss": 0.027745448225544075, "val_loss": 0.027823254466056824, "global_step": 712 } ], "is_current": false, "weights_repository": "historical diagnostic package; keep separate from the final 128-episode adapter repo" }, { "id": "xperience10m_qwen3_omni_128ep_structured_json_v3_strict_label_prompt_reuse_lora_eval_test_full", "title": "Qwen3-Omni LoRA", "status": "verified", "backbone": "qwen3_omni_lora", "dataset_contract": "xperience10m_episode_json_qa_v1", "training_objective": "structured_episode_understanding_json_qa", "source": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v3_strict_label_prompt_reuse_lora_eval_test_full/verified_result_summary.json", "dataset_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605", "train_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora", "eval_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v3_strict_label_prompt_reuse_lora_eval_test_full", "counts": { "dataset_samples": 3808, "dataset_episodes": 119, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "train_samples": 2848, "val_samples": 512, "eval_samples": 448, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "json_validity_rate": 1.0, "action_macro_f1": 0.0021983997167007384, "subtask_accuracy": 0.002232142857142857, "transition_accuracy": 0.9732142857142857, "next_action_accuracy": 0.03125, "contact_accuracy": 0.7209821428571429, "object_micro_f1": 0.30688228657389993, "held_out_episode_count": 14 }, "history": [ { "epoch": 1, "train_loss": 0.41282760031950355, "val_loss": 0.03288277983665466, "global_step": 356 }, { "epoch": 2, "train_loss": 0.027745448225544075, "val_loss": 0.027823254466056824, "global_step": 712 } ], "is_current": false, "weights_repository": "historical diagnostic package; keep separate from the final 128-episode adapter repo" }, { "id": "xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora_eval_test_full", "title": "Qwen3-Omni LoRA", "status": "verified", "backbone": "qwen3_omni_lora", "dataset_contract": "xperience10m_episode_json_qa_v1", "training_objective": "structured_episode_understanding_json_qa", "source": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora_eval_test_full/verified_result_summary.json", "dataset_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605", "train_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora", "eval_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora_eval_test_full", "counts": { "dataset_samples": 3808, "dataset_episodes": 119, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "train_samples": 2848, "val_samples": 512, "eval_samples": 448, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "json_validity_rate": 1.0, "action_macro_f1": 0.0018678269676001454, "subtask_accuracy": 0.0, "transition_accuracy": 0.9732142857142857, "next_action_accuracy": 0.033482142857142856, "contact_accuracy": 0.7299107142857143, "object_micro_f1": 0.31099781500364165, "held_out_episode_count": 14 }, "history": [ { "epoch": 1, "train_loss": 0.40796751019628613, "val_loss": 0.03258896619081497, "global_step": 356 }, { "epoch": 2, "train_loss": 0.027628723937453012, "val_loss": 0.027754632756114006, "global_step": 712 }, { "epoch": 3, "train_loss": 0.02446955946807781, "val_loss": 0.026343274861574173, "global_step": 1068 }, { "epoch": 4, "train_loss": 0.022728607045444712, "val_loss": 0.025629229843616486, "global_step": 1424 } ], "is_current": false, "weights_repository": "historical diagnostic package; keep separate from the final 128-episode adapter repo" } ], "interpretation": "This layer contains the held-out foundation-model packages. Qwen3-Omni packages evaluate structured JSON task prediction; Cosmos3-Nano evaluates a future-window world-model compatibility adapter; Cosmos3-Super Reasoner evaluates staged base weights through vLLM on the JSON task; Cosmos3-Super Forward-Dynamics LoRA is the first Super adapter branch and evaluates camera-pose-conditioned future vision velocity loss." } ], "model_groups": [ { "id": "task_head_baselines", "model_family": "Minimal and Neural Task Heads", "model_type": "lightweight supervised/self-supervised task heads", "weight_repository": "https://huggingface.co/cy0307/ropedia-xperience-10m-task-baselines", "one_episode_runs": [ { "id": "task_heads_single_episode_public_sample", "title": "Single-Episode Public-Sample Task Suite", "scope": "one public Xperience-10M sample episode", "status": "verified", "source": "results/episode_task_suite/summary_report.json", "split": "chronological 70/30 within one episode", "counts": { "episodes": 1, "windows": 1161, "frames": 5821, "feature_dim": 8546, "task_count": 12, "neural_task_count": 12 }, "weights": "baseline model files in the baseline model repo; no foundation-model weights", "interpretation": "Raw multimodal feature task harness on the public sample." } ], "multi_episode_128_runs": [ { "id": "task_heads_128_episode_metadata_baselines", "title": "128-Episode Aligned Simple/NN Baselines", "scope": "selected 128-episode 96/16/16 split", "status": "pass", "source": "results/omni_finetune/multi_episode_128_task_baselines/BASELINE_ALIGNMENT_REPORT.md", "split": "train/val/test by selected episode/session", "counts": { "rows": 3808, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "episode_counts": { "test": 16, "train": 96, "val": 16 }, "task_count": 12, "simple_supported_task_count": 8, "neural_supported_task_count": 6 }, "weights": "metadata/text baseline artifacts; raw 128 sensor-feature model weights not yet complete", "interpretation": "Same selected 96/16/16 split and task ids as the model branches, but metadata/text features only." } ], "comparison_note": "This is the cleanest 1-episode versus 128-episode grouping for the same simple/NN task-head family, but the feature surface changes from raw public-sample features to public-safe 128-episode metadata/text features." }, { "id": "qwen3_omni_lora", "model_family": "Qwen3-Omni LoRA", "model_type": "PEFT LoRA adapter over Qwen/Qwen3-Omni-30B-A3B-Instruct", "weight_repository": "https://huggingface.co/cy0307/ropedia-qwen3-omni-lora-128ep", "one_episode_runs": [ { "id": "qwen3_omni_sensor_adapter_smoke_1ep", "title": "Qwen3-Omni Sensor-Adapter Smoke", "scope": "one public Xperience-10M sample episode", "status": "verified_smoke", "source": "results/omni_exploration/qwen3_adapter_smoke/metrics.json", "split": "single_episode_chronological", "counts": { "episodes": 1, "windows": 59, "train_windows": 41, "test_windows": 18, "feature_dim": 4262, "adapter_tokens": 11 }, "primary_metrics": { "accuracy": 0.0, "macro_f1": 0.0, "train_final_loss": 1.4479121318677577 }, "base_model_target": "Qwen/Qwen3-Omni-30B-A3B-Thinking", "qwen3_loaded": false, "weights": "no Qwen3 base weights or LoRA adapter weights; adapter-token readiness smoke only", "interpretation": "This validates the sensor-adapter token path on one real episode before loading or LoRA-tuning Qwen3-Omni. It is not comparable to the 128-episode held-out LoRA result." } ], "readiness_runs": [ { "id": "xperience10m_qwen3_omni_128ep_fullparam_smoke_preemptible_8gpu_20260609", "title": "Full-Parameter 1-Step Feasibility Smoke", "scope_label": "full-param gate", "scope": "1 optimizer step over 8 train samples", "status": "passed", "source": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_smoke_preemptible_8gpu_20260609/fullparam_feasibility_summary.json", "split": "selected 128-episode train split", "counts": { "samples": 8, "steps": 1, "num_processes": 8 }, "primary_metrics": { "full_parameter_gate": "passed", "observed_train_steps": 1, "final_step_loss": 1.2726006507873535, "epoch_train_loss": 1.2726006507873535, "checkpoint_saved": false }, "weights": "no full-parameter checkpoint or public weights; save_mode=none", "interpretation": "Full-parameter FSDP feasibility evidence only. This gate is not a held-out model result, full fine-tune, checkpoint release, or public weight package." }, { "id": "xperience10m_qwen3_omni_128ep_fullparam_shorttrain8_preemptible_8gpu_20260609", "title": "Full-Parameter 8-Step Short Train", "scope_label": "full-param gate", "scope": "8 optimizer steps over 64 train samples", "status": "passed", "source": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_shorttrain8_preemptible_8gpu_20260609/fullparam_shorttrain8_summary.json", "split": "selected 128-episode train split", "counts": { "samples": 64, "steps": 8, "num_processes": 8 }, "primary_metrics": { "full_parameter_gate": "passed", "observed_train_steps": 8, "final_step_loss": 1.180522084236145, "epoch_train_loss": 1.2190196067094803, "checkpoint_saved": false }, "weights": "no full-parameter checkpoint or public weights; save_mode=none", "interpretation": "Full-parameter FSDP feasibility evidence only. This gate is not a held-out model result, full fine-tune, checkpoint release, or public weight package." }, { "id": "xperience10m_qwen3_omni_128ep_fullparam_pilot32_preemptible_8gpu_20260609", "title": "Full-Parameter 32-Step Pilot", "scope_label": "full-param gate", "scope": "32 optimizer steps over 256 train samples", "status": "passed", "source": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_pilot32_preemptible_8gpu_20260609/fullparam_pilot32_summary.json", "split": "selected 128-episode train split", "counts": { "samples": 256, "steps": 32, "num_processes": 8 }, "primary_metrics": { "full_parameter_gate": "passed", "observed_train_steps": 32, "final_step_loss": 0.2206273376941681, "epoch_train_loss": 0.8451133379712701, "checkpoint_saved": false }, "weights": "no full-parameter checkpoint or public weights; save_mode=none", "interpretation": "Full-parameter FSDP feasibility evidence only. This gate is not a held-out model result, full fine-tune, checkpoint release, or public weight package." }, { "id": "xperience10m_qwen3_omni_128ep_fullparam_pilot64_preemptible_8gpu_20260609", "title": "Full-Parameter 64-Step Pilot", "scope_label": "full-param gate", "scope": "64 optimizer steps over 512 train samples", "status": "passed", "source": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_pilot64_preemptible_8gpu_20260609/fullparam_pilot64_summary.json", "split": "selected 128-episode train split", "counts": { "samples": 512, "steps": 64, "num_processes": 8 }, "primary_metrics": { "full_parameter_gate": "passed", "observed_train_steps": 64, "final_step_loss": 0.011219973675906658, "epoch_train_loss": 0.4434075650788145, "checkpoint_saved": false }, "weights": "no full-parameter checkpoint or public weights; save_mode=none", "interpretation": "Full-parameter FSDP feasibility evidence only. This gate is not a held-out model result, full fine-tune, checkpoint release, or public weight package." }, { "id": "xperience10m_qwen3_omni_128ep_fullparam_pilot128_preemptible_8gpu_20260609", "title": "Full-Parameter 128-Step Opportunistic Pilot", "scope_label": "full-param gate", "scope": "planned 128 optimizer steps over 1024 train samples; preempted for Qwen v5 handoff", "status": "preempted_for_qwen_v5_handoff", "source": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_pilot128_preemptible_8gpu_20260609/fullparam_pilot128_summary.json", "split": "selected 128-episode train split", "counts": { "samples": 1024, "steps": 0, "num_processes": 8 }, "primary_metrics": { "full_parameter_gate": "preempted_for_qwen_v5_handoff", "observed_train_steps": 0, "final_step_loss": null, "epoch_train_loss": null, "checkpoint_saved": false }, "weights": "no full-parameter checkpoint or public weights; save_mode=none", "interpretation": "Full-parameter FSDP feasibility evidence only. This gate is not a held-out model result, full fine-tune, checkpoint release, or public weight package." }, { "id": "xperience10m_qwen3_omni_128ep_fullparam_pilot128_after_qwen_v5_preemptible_8gpu_20260609", "title": "Full-Parameter 128-Step Post-Qwen-v5 Pilot", "scope_label": "full-param gate", "scope": "128 optimizer steps over 1024 train samples after verified Qwen v5 handoff", "status": "passed", "source": "results/omni_finetune/xperience10m_qwen3_omni_128ep_fullparam_pilot128_after_qwen_v5_preemptible_8gpu_20260609/training_metadata.json", "split": "selected 128-episode train split", "counts": { "samples": 1024, "steps": 128, "num_processes": 8 }, "primary_metrics": { "full_parameter_gate": "passed", "observed_train_steps": 128, "final_step_loss": 0.0136940386146307, "epoch_train_loss": 0.21579630990163423, "checkpoint_saved": false }, "weights": "no full-parameter checkpoint or public weights; save_mode=none", "interpretation": "Full-parameter FSDP feasibility evidence only. This gate is not a held-out model result, full fine-tune, checkpoint release, or public weight package." } ], "multi_episode_128_runs": [ { "id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_eval", "title": "Qwen3-Omni LoRA", "status": "verified", "backbone": "qwen3_omni_lora", "dataset_contract": "xperience10m_episode_json_qa_v1", "training_objective": "structured_episode_understanding_json_qa", "source": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_eval/verified_result_summary.json", "dataset_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605", "train_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_lora", "eval_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_eval", "counts": { "dataset_samples": 3808, "dataset_episodes": 119, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "train_samples": 2848, "val_samples": 512, "eval_samples": 448, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "json_validity_rate": 0.875, "action_macro_f1": 0.0026621494447581404, "subtask_accuracy": 0.006696428571428571, "transition_accuracy": 0.8504464285714286, "next_action_accuracy": 0.024553571428571428, "contact_accuracy": 0.6450892857142857, "object_micro_f1": 0.22299431459254582, "held_out_episode_count": 14 }, "history": [ { "epoch": 1, "train_loss": 0.41304643672440994, "val_loss": 0.0330660454928875, "global_step": 356 } ], "is_current": false, "weights_repository": "historical diagnostic package; keep separate from the final 128-episode adapter repo" }, { "id": "xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu_lora_fsdp_full_train_noval_tail_logits_fullstatesave_v6_eval_test_full", "title": "Qwen3-Omni LoRA", "status": "verified", "backbone": "qwen3_omni_lora", "dataset_contract": "xperience10m_episode_json_qa_v1", "training_objective": "structured_episode_understanding_json_qa", "source": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu_lora_fsdp_full_train_noval_tail_logits_fullstatesave_v6_eval_test_full/verified_result_summary.json", "dataset_run_id": "xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu", "train_run_id": "xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu_lora_fsdp_full_train_noval_tail_logits_fullstatesave_v6", "eval_run_id": "xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu_lora_fsdp_full_train_noval_tail_logits_fullstatesave_v6_eval_test_full", "counts": { "dataset_samples": 3808, "dataset_episodes": 119, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "train_samples": 2848, "val_samples": 0, "eval_samples": 448, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "json_validity_rate": 0.8526785714285714, "action_macro_f1": 0.00213753459655099, "subtask_accuracy": 0.004464285714285714, "transition_accuracy": 0.828125, "next_action_accuracy": 0.022321428571428572, "contact_accuracy": 0.6517857142857143, "object_micro_f1": 0.23062730627306272, "held_out_episode_count": 14 }, "history": [ { "epoch": 1, "train_loss": 0.4121775626560694, "val_loss": null, "global_step": 356 } ], "is_current": false, "weights_repository": "historical diagnostic package; keep separate from the final 128-episode adapter repo" }, { "id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora_eval_test_full", "title": "Qwen3-Omni LoRA", "status": "verified", "backbone": "qwen3_omni_lora", "dataset_contract": "xperience10m_episode_json_qa_v1", "training_objective": "structured_episode_understanding_json_qa", "source": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora_eval_test_full/verified_result_summary.json", "dataset_run_id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora", "train_run_id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora", "eval_run_id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora_eval_test_full", "counts": { "dataset_samples": 34269, "dataset_episodes": 119, "split_counts": { "test": 4032, "train": 25629, "val": 4608 }, "train_samples": 25629, "val_samples": 1024, "eval_samples": 4032, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "json_validity_rate": 1.0, "action_macro_f1": 0.002289711036077459, "subtask_accuracy": 0.011194029850746268, "transition_accuracy": 0.9908234126984127, "next_action_accuracy": 0.053618594823032224, "contact_accuracy": 0.7864583333333334, "object_micro_f1": 0.31614599936244814, "held_out_episode_count": 14 }, "history": [ { "epoch": 1, "train_loss": 0.06255286606544624, "val_loss": 0.02668904885649681, "global_step": 3204 } ], "is_current": true, "weights_repository": "https://huggingface.co/cy0307/ropedia-qwen3-omni-lora-128ep" }, { "id": "xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora_eval_test_full", "title": "Qwen3-Omni LoRA", "status": "verified", "backbone": "qwen3_omni_lora", "dataset_contract": "xperience10m_episode_json_qa_v1", "training_objective": "structured_episode_understanding_json_qa", "source": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora_eval_test_full/verified_result_summary.json", "dataset_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605", "train_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora", "eval_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora_eval_test_full", "counts": { "dataset_samples": 3808, "dataset_episodes": 119, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "train_samples": 2848, "val_samples": 512, "eval_samples": 448, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "json_validity_rate": 0.9977678571428571, "action_macro_f1": 0.0024331644885523347, "subtask_accuracy": 0.002232142857142857, "transition_accuracy": 0.9709821428571429, "next_action_accuracy": 0.029017857142857144, "contact_accuracy": 0.71875, "object_micro_f1": 0.30160427807486634, "held_out_episode_count": 14 }, "history": [ { "epoch": 1, "train_loss": 0.41282760031950355, "val_loss": 0.03288277983665466, "global_step": 356 }, { "epoch": 2, "train_loss": 0.027745448225544075, "val_loss": 0.027823254466056824, "global_step": 712 } ], "is_current": false, "weights_repository": "historical diagnostic package; keep separate from the final 128-episode adapter repo" }, { "id": "xperience10m_qwen3_omni_128ep_structured_json_v3_strict_label_prompt_reuse_lora_eval_test_full", "title": "Qwen3-Omni LoRA", "status": "verified", "backbone": "qwen3_omni_lora", "dataset_contract": "xperience10m_episode_json_qa_v1", "training_objective": "structured_episode_understanding_json_qa", "source": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v3_strict_label_prompt_reuse_lora_eval_test_full/verified_result_summary.json", "dataset_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605", "train_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora", "eval_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v3_strict_label_prompt_reuse_lora_eval_test_full", "counts": { "dataset_samples": 3808, "dataset_episodes": 119, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "train_samples": 2848, "val_samples": 512, "eval_samples": 448, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "json_validity_rate": 1.0, "action_macro_f1": 0.0021983997167007384, "subtask_accuracy": 0.002232142857142857, "transition_accuracy": 0.9732142857142857, "next_action_accuracy": 0.03125, "contact_accuracy": 0.7209821428571429, "object_micro_f1": 0.30688228657389993, "held_out_episode_count": 14 }, "history": [ { "epoch": 1, "train_loss": 0.41282760031950355, "val_loss": 0.03288277983665466, "global_step": 356 }, { "epoch": 2, "train_loss": 0.027745448225544075, "val_loss": 0.027823254466056824, "global_step": 712 } ], "is_current": false, "weights_repository": "historical diagnostic package; keep separate from the final 128-episode adapter repo" }, { "id": "xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora_eval_test_full", "title": "Qwen3-Omni LoRA", "status": "verified", "backbone": "qwen3_omni_lora", "dataset_contract": "xperience10m_episode_json_qa_v1", "training_objective": "structured_episode_understanding_json_qa", "source": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora_eval_test_full/verified_result_summary.json", "dataset_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605", "train_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora", "eval_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora_eval_test_full", "counts": { "dataset_samples": 3808, "dataset_episodes": 119, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "train_samples": 2848, "val_samples": 512, "eval_samples": 448, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "json_validity_rate": 1.0, "action_macro_f1": 0.0018678269676001454, "subtask_accuracy": 0.0, "transition_accuracy": 0.9732142857142857, "next_action_accuracy": 0.033482142857142856, "contact_accuracy": 0.7299107142857143, "object_micro_f1": 0.31099781500364165, "held_out_episode_count": 14 }, "history": [ { "epoch": 1, "train_loss": 0.40796751019628613, "val_loss": 0.03258896619081497, "global_step": 356 }, { "epoch": 2, "train_loss": 0.027628723937453012, "val_loss": 0.027754632756114006, "global_step": 712 }, { "epoch": 3, "train_loss": 0.02446955946807781, "val_loss": 0.026343274861574173, "global_step": 1068 }, { "epoch": 4, "train_loss": 0.022728607045444712, "val_loss": 0.025629229843616486, "global_step": 1424 } ], "is_current": false, "weights_repository": "historical diagnostic package; keep separate from the final 128-episode adapter repo" } ], "comparison_note": "The one-episode Qwen entry is only a sensor-adapter smoke test with Qwen3 weights unloaded. The 128-episode entries are real held-out LoRA diagnostics; the current final adapter belongs in the separate Qwen model repo. The full-parameter rows are feasibility gates only and intentionally publish no checkpoints or full-parameter weights." }, { "id": "cosmos3_nano_world_model", "model_family": "Cosmos3-Nano Future-Window World Model", "model_type": "world-model/future-window branch", "weight_repository": "planned: cy0307/ropedia-cosmos3-nano-future-window-lora-128ep after real adapter weights exist", "one_episode_runs": [ { "id": "cosmos3_nano_one_episode", "title": "Cosmos3-Nano One-Episode Fine-Tune", "scope": "one public Xperience-10M sample episode", "status": "not_run", "source": null, "weights": "none", "interpretation": "No Cosmos3 one-episode adapter or diffusion-weight fine-tune is currently published. Use the public-sample task suite only as model-agnostic evidence." } ], "multi_episode_128_runs": [ { "id": "xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full", "title": "Cosmos3-Nano Future-Window World Model", "status": "verified", "backbone": "cosmos_world_model", "dataset_contract": "xperience10m_future_window_world_model_v0", "training_objective": "future_window_and_action_conditioned_world_modeling", "source": "results/omni_finetune/verified_public/xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full/verified_result_summary.json", "dataset_run_id": "xperience10m_cosmos3_nano_128ep_future_window_h5_compat", "train_run_id": "xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter", "eval_run_id": "xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full", "counts": { "dataset_samples": 3213, "dataset_episodes": 119, "split_counts": { "train": 2403, "test": 378, "val": 432 }, "train_samples": 2403, "val_samples": 432, "eval_samples": 378, "held_out_episode_count": 14, "num_processes": 1 }, "primary_metrics": { "future_retrieval_mrr": 0.022138720585222767, "future_retrieval_recall_at_5": 0.015873015873015872, "temporal_consistency": 0.09523809523809523, "feature_reconstruction_error": 3479.218317102503, "transition_accuracy": 0.9682539682539683, "contact_accuracy": 0.7433862433862434, "held_out_episode_count": 14 }, "history": [ { "epoch": 0, "train_loss": null, "val_loss": null, "note": "closed-form mean-delta adapter; no Cosmos diffusion weights fine-tuned in this compatibility run" } ], "is_current": true, "weights_repository": "planned separate Cosmos3 model repo after a real Cosmos diffusion/LoRA fine-tune exists; current result remains artifacts-only" } ], "comparison_note": "The current 128-episode Cosmos result is a public-safe future-window compatibility adapter. It is not yet a full Cosmos diffusion/LoRA weight release." }, { "id": "cosmos3_super_reasoner", "model_family": "Cosmos3-Super Reasoner", "model_type": "base-weight vLLM Reasoner evaluation over nv-community/Cosmos3-Super", "weight_repository": "none for this run; staged base weights only, no new fine-tuned weights", "one_episode_runs": [ { "id": "cosmos3_super_one_episode", "title": "Cosmos3-Super One-Episode Fine-Tune", "scope": "one public Xperience-10M sample episode", "status": "not_run", "source": null, "weights": "none", "interpretation": "No one-episode Cosmos3-Super adapter or fine-tuned weight run is published. The available Super result is the 128-episode held-out base-weight evaluation." } ], "readiness_runs": [ { "id": "xperience10m_cosmos3_super_training_readiness_20260607", "title": "Cosmos3-Super Training Readiness Probe", "scope": "selected 128-episode 96/16/16 JSON-task dataset and staged Cosmos3-Super runtime", "status": "blocked_until_trainer_implemented", "source": "results/omni_finetune/xperience10m_cosmos3_super_training_readiness_20260607/training_readiness.json", "split": "train/val/test by selected episode/session", "counts": { "dataset_samples": 3808, "split_counts": { "test": { "samples": 448, "episodes": 14, "actions": 189 }, "train": { "samples": 2848, "episodes": 89, "actions": 885 }, "val": { "samples": 512, "episodes": 16, "actions": 223 } } }, "primary_metrics": { "diffusers_runtime_supported": true, "chat_sft_supported": false, "weights_updated": false }, "weights": "none; readiness audit only, no adapter checkpoint", "interpretation": "This probe confirms the staged Cosmos3-Super Diffusers/GPU runtime and the same JSON QA dataset are visible. It predates the camera-pose action-target export, so use the 20260608 contract audit for the current trainer-readiness status." }, { "id": "xperience10m_cosmos3_super_training_readiness_metadata_a100_20260609", "title": "Cosmos3-Super Remote Staging Readiness Probe", "scope_label": "staging readiness", "scope": "secondary 4-GPU staging tree, JSON-task dataset visibility, and metadata-only Cosmos3-Super runtime probe", "status": "blocked_until_trainer_implemented", "source": "results/omni_finetune/xperience10m_cosmos3_super_training_readiness_metadata_a100_20260609/training_readiness.json", "split": "train/val/test by selected episode/session", "counts": { "dataset_samples": 3808, "split_counts": { "test": { "samples": 448, "episodes": 14, "actions": 189 }, "train": { "samples": 2848, "episodes": 89, "actions": 885 }, "val": { "samples": 512, "episodes": 16, "actions": 223 } } }, "primary_metrics": { "model_files_visible": false, "diffusers_runtime_supported": false, "cuda_device_count": 4, "weights_updated": false }, "weights": "none; staging readiness audit only, no adapter checkpoint", "interpretation": "This metadata-only probe checks the secondary 4-GPU staging tree without loading the model pipeline or updating weights. It confirms the JSON task dataset is present, but the Cosmos3-Super model files and Diffusers runtime are not staged there yet, so real Super training should wait for model/runtime staging or run on the already prepared main host." }, { "id": "xperience10m_cosmos3_super_training_contract_audit_camera_pose_20260608", "title": "Cosmos3-Super Camera-Pose Target Audit", "scope_label": "action target contract", "scope": "selected 128-episode 96/16/16 dataset augmented with camera_pose proxy cosmos_action_target records", "status": "ready_for_forward_dynamics_trainer", "source": "results/omni_finetune/xperience10m_cosmos3_super_training_contract_audit_camera_pose_20260608/training_contract_audit.json", "split": "train/val/test by selected episode/session", "counts": { "dataset_samples": 3808, "rows_with_action_target": 3808, "valid_action_targets": 3808, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "episode_split_counts": { "test": 14, "train": 89, "val": 16 } }, "primary_metrics": { "domain_name": "camera_pose", "raw_action_dim": 9, "mode": "forward_dynamics", "valid_action_targets": 3808, "weights_updated": false }, "weights": "none; action-target contract audit only, no adapter checkpoint", "interpretation": "The selected dataset now has valid Cosmos3 camera_pose forward_dynamics targets for an egocentric camera-motion proxy. These remove the target-schema blocker for action-conditioned world-model training, but they supervise noisy vision tokens rather than preds_action. The remaining work is a trainable Cosmos3-Super implementation that can backpropagate through this loss surface at the required memory scale; action-token prediction needs a separate policy or inverse-dynamics target export." }, { "id": "xperience10m_cosmos3_super_action_packer_schema_smoke_20260608", "title": "Cosmos3-Super Action Batch Packer Smoke", "scope_label": "batch packer", "scope": "one selected train row from the camera_pose forward_dynamics augmented JSONL", "status": "pass", "source": "results/omni_finetune/xperience10m_cosmos3_super_action_packer_schema_smoke_20260608/packer_summary.json", "split": "train", "counts": { "samples": 1, "raw_action_rows": 8, "raw_action_dim": 9 }, "primary_metrics": { "mode": "forward_dynamics", "loss_surface": "vision_velocity_conditioned_on_camera_pose", "pipeline_loaded": false, "weights_updated": false }, "weights": "none; schema-only packer smoke, no adapter checkpoint", "interpretation": "The selected row maps to a camera_pose forward_dynamics contract. In the installed Cosmos3 pipeline this uses raw actions as conditioning and supervises noisy vision tokens; it does not supervise preds_action." } ], "multi_episode_128_runs": [ { "id": "xperience10m_cosmos3_super_reasoner_128ep_test_full_20260607", "title": "Cosmos3-Super Reasoner", "status": "verified", "backbone": "cosmos3_super_reasoner", "dataset_contract": "xperience10m_episode_json_qa_v1", "training_objective": "zero_shot_structured_episode_understanding_json_qa_via_vllm_reasoner", "source": "results/omni_finetune/verified_public/xperience10m_cosmos3_super_reasoner_128ep_test_full_20260607/verified_result_summary.json", "dataset_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605", "train_run_id": "xperience10m_cosmos3_super_reasoner_base_vllm_8gpu_20260607", "eval_run_id": "xperience10m_cosmos3_super_reasoner_128ep_test_full_20260607", "counts": { "dataset_samples": 3808, "dataset_episodes": 119, "split_counts": { "train": 2848, "val": 512, "test": 448 }, "train_samples": 2848, "val_samples": 512, "eval_samples": 448, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "json_validity_rate": 0.5111607142857143, "action_macro_f1": 0.0008284021201089245, "subtask_accuracy": 0.0, "transition_accuracy": 0.36830357142857145, "next_action_accuracy": 0.013392857142857142, "contact_accuracy": 0.32142857142857145, "object_micro_f1": 0.13704276146316333, "held_out_episode_count": 14 }, "history": [], "is_current": true, "weights_repository": "none for this run: staged base nv-community/Cosmos3-Super weights were evaluated through vLLM; create a separate repo only after new adapter or fine-tuned weights exist" } ], "comparison_note": "Cosmos3-Super is now represented by a verified 448-window held-out Reasoner evaluation on the same JSON task as Qwen3. It uses staged base weights through vLLM, so it is a model-branch diagnostic, not a weight release. A camera-pose proxy forward-dynamics target export now passes the contract audit and schema-only packer smoke; the separate Forward-Dynamics LoRA group records the trainable adapter run and loss-based held-out evaluation." }, { "id": "cosmos3_super_forward_dynamics", "model_family": "Cosmos3-Super Forward-Dynamics LoRA", "model_type": "PEFT LoRA over nv-community/Cosmos3-Super for camera-pose-conditioned future vision velocity", "weight_repository": "https://huggingface.co/cy0307/ropedia-cosmos3-super-forward-dynamics-lora-128ep", "one_episode_runs": [ { "id": "cosmos3_super_forward_dynamics_overfit_smoke", "title": "Cosmos3-Super Forward-Dynamics Overfit Smoke", "scope": "small overfit smoke before 128-episode scale-up", "status": "verified_smoke", "source": "results/omni_finetune/xperience10m_cosmos3_super_forward_dynamics_lora_overfit_after_qwen_v4_20260608_fsdp8_attn256_gradfix_savefix2/", "weights": "local repaired LoRA smoke adapter, not public packaged as final", "interpretation": "Validated the trainable adapter path, FSDP save repair, and Diffusers load before the full 128-episode run." } ], "multi_episode_128_runs": [ { "id": "xperience10m_cosmos3_super_forward_dynamics_lora_128ep_train1epoch_256_attn_full8gpu_20260608_eval_test_full_fsdp", "title": "Cosmos3-Super Forward-Dynamics LoRA", "status": "verified", "backbone": "cosmos3_super_forward_dynamics", "dataset_contract": "xperience10m_camera_pose_forward_dynamics_v1", "training_objective": "camera_pose_conditioned_future_vision_velocity_lora", "source": "results/omni_finetune/verified_public/xperience10m_cosmos3_super_forward_dynamics_lora_128ep_train1epoch_256_attn_full8gpu_20260608_eval_test_full_fsdp/verified_result_summary.json", "dataset_run_id": "xperience10m_cosmos3_camera_pose_targets_20260608", "train_run_id": "xperience10m_cosmos3_super_forward_dynamics_lora_128ep_train1epoch_256_attn_full8gpu_20260608", "eval_run_id": "xperience10m_cosmos3_super_forward_dynamics_lora_128ep_train1epoch_256_attn_full8gpu_20260608_eval_test_full_fsdp", "counts": { "dataset_samples": 3808, "dataset_episodes": 119, "split_counts": { "test": 448, "train": 2848, "val": 512 }, "train_samples": 2848, "val_samples": 512, "eval_samples": 448, "held_out_episode_count": 14, "num_processes": 8 }, "primary_metrics": { "adapter_parameter_numel": 26214400, "held_out_episode_count": 14, "test_forward_dynamics_mse": 3.6853174321087345, "train_final_loss": 1.0785235166549683, "val_forward_dynamics_mse": 4.008244896889664 }, "history": [ { "epoch": 1, "note": "FSDP 8-GPU LoRA over camera-pose-conditioned future vision velocity loss; adapter weights are excluded from this public package.", "train_loss": 1.0785235166549683, "val_loss": 4.008244896889664 } ], "is_current": true, "weights_repository": "https://huggingface.co/cy0307/ropedia-cosmos3-super-forward-dynamics-lora-128ep" } ], "comparison_note": "This is the first verified Cosmos3-Super fine-tuned adapter branch. Its metric is forward-dynamics MSE, so compare it to world-model loss or future-prediction targets, not to Qwen JSON classification accuracy." } ], "model_group_reading_notes": [ "Use model_groups when comparing one-episode and 128-episode artifacts within the same model family.", "Task-head baselines have both a one-episode public-sample run and a 128-episode same-split metadata/text run.", "Qwen3-Omni has a one-episode sensor-adapter smoke test, full-parameter feasibility gates, and separate 128-episode LoRA diagnostic packages; the newest verified full-eval 128-episode adapter belongs in the Qwen LoRA model repo.", "Cosmos3-Nano has a 128-episode future-window compatibility package.", "Cosmos3-Super now has both a 128-episode base-weight Reasoner evaluation on the JSON task and a fine-tuned forward-dynamics LoRA branch over camera-pose proxy targets." ], "pending": [ "Use the verified Qwen3 v5 dense multiscale full-eval package as the current Qwen row; older Qwen package rows remain historical diagnostics for comparison." ] }