| { |
| "title": "Verified Qwen3-Omni LoRA 128-Episode Held-Out Result", |
| "status": "verified_latest_qwen3_v6_diagnostic_result", |
| "status_date": "2026-06-14", |
| "backbone": "Qwen/Qwen3-Omni-30B-A3B-Instruct", |
| "adapter": "Qwen3-Omni LoRA", |
| "dataset": "Ropedia Xperience-10M selected 128-episode pilot", |
| "split_policy": { |
| "unit": "episode", |
| "selected_episode_counts": { |
| "train": 96, |
| "val": 16, |
| "test": 16 |
| }, |
| "exported_window_counts": { |
| "train": 25629, |
| "val": 4608, |
| "test": 4032 |
| }, |
| "exported_episode_counts": { |
| "train": 89, |
| "val": 16, |
| "test": 14 |
| }, |
| "skipped_selected_episodes": 9, |
| "leakage_policy": "Train, validation, and test are separated by episode/session; test windows are used only for held-out evaluation." |
| }, |
| "training": { |
| "num_processes": 8, |
| "epochs": 2, |
| "lora_rank": 64, |
| "lora_alpha": 128, |
| "lora_dropout": 0.05, |
| "learning_rate": 0.00005, |
| "num_train_samples": 25629, |
| "num_val_samples": 2048, |
| "history": [ |
| { |
| "epoch": 1, |
| "train_loss": 0.05208605339353295, |
| "val_loss": 0.026512427255511284, |
| "global_step": 3204 |
| }, |
| { |
| "epoch": 2, |
| "train_loss": 0.013760763933660042, |
| "val_loss": 0.032345958054065704, |
| "global_step": 6408 |
| } |
| ], |
| "loss": "answer-token cross entropy over supervised JSON tokens", |
| "note": "This current Qwen3-Omni LoRA result is the v6 rank64/lr5e-5 dense multiscale held-out evaluation on the selected 96/16/16 episode setup." |
| }, |
| "evaluation": { |
| "split": "test", |
| "num_samples": 4032, |
| "held_out_episode_count": 14, |
| "json_validity_rate": 0.9990079365079365, |
| "action_macro_f1": 0.0028830723979596335, |
| "subtask_accuracy": 0.0037313432835820895, |
| "transition_accuracy": 0.9898313492063492, |
| "next_action_accuracy": 0.04305335446381405, |
| "contact_accuracy": 0.8177083333333334, |
| "object_micro_f1": 0.3064982378331287, |
| "quality_target": { |
| "json_validity_rate": 0.98, |
| "status": "met" |
| }, |
| "previous_v5_json_validity_rate": 1.0, |
| "previous_v5_action_macro_f1": 0.002289711036077459, |
| "previous_v5_subtask_accuracy": 0.011194029850746268, |
| "previous_v5_next_action_accuracy": 0.053618594823032224, |
| "previous_v5_contact_accuracy": 0.7864583333333334, |
| "previous_v5_object_micro_f1": 0.31614599936244814 |
| }, |
| "interpretation": "This is the latest verified Qwen3-Omni LoRA diagnostic result for the selected 128-episode setup. The v6 rank64/lr5e-5 package keeps JSON validity above the 98% target and improves action macro-F1 and contact accuracy versus the pinned v5 release row, but slightly regresses JSON validity, subtask accuracy, next-action accuracy, transition accuracy, and object micro-F1. Treat it as the latest diagnostic branch, not as a globally stronger replacement for v5.", |
| "public_package": { |
| "path": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora_eval_test_full", |
| "audit_status": "pass", |
| "contains_raw_xperience10m_data": false, |
| "contains_qwen_base_weights": false, |
| "contains_lora_weights": false, |
| "adapter_weights_repo": "cy0307/ropedia-qwen3-omni-lora-128ep" |
| }, |
| "release_policy": { |
| "latest_verified_qwen_row": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora_eval_test_full", |
| "pinned_release_tag": "ropedia-xperience-10m-v5", |
| "pinned_release_reason": "v5 remains the prior stable release tag; v6 is published on main/HF as the latest verified branch and can receive a separate v6 release tag." |
| }, |
| "required_next_steps": [ |
| "Use results/omni_finetune/QWEN3_V5_V6_COMPARISON_20260614.md before deciding whether v6 should become a formal release tag.", |
| "Use the v6 predictions for action/contact error analysis, and compare v5 for subtask, next-action, and object regressions.", |
| "Keep full-parameter Qwen runs as feasibility gates until there is a storage plan for checkpoints or mergeable full-weight deltas.", |
| "Use the verified Cosmos3-Super Forward-Dynamics LoRA package as a separate world-model branch: it updates adapter weights over camera-pose proxy future-vision-velocity targets, not Qwen-style JSON action labels." |
| ] |
| } |
|
|
