| { |
| "title": "Qwen3-Omni v5 versus v6 verified comparison", |
| "status": "pass", |
| "generated_at_utc": "2026-06-14T00:00:00+00:00", |
| "comparison_scope": "Verified Qwen3-Omni LoRA held-out test packages on the same dense multiscale selected 128-episode dataset.", |
| "release_policy": { |
| "latest_verified_qwen_row": "v6", |
| "pinned_release_tag": "ropedia-xperience-10m-v5", |
| "recommendation": "Publish v6 as the latest verified branch and create a separate v6 tag only if the project wants a formal experimental release; do not move the v5 tag." |
| }, |
| "runs": { |
| "v5": { |
| "eval_run_id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora_eval_test_full", |
| "train_run_id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora", |
| "epochs": 1, |
| "eval_samples": 4032, |
| "held_out_episode_count": 14, |
| "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 |
| } |
| }, |
| "v6": { |
| "eval_run_id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora_eval_test_full", |
| "train_run_id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora", |
| "epochs": 2, |
| "lora_rank": 64, |
| "learning_rate": 0.00005, |
| "eval_samples": 4032, |
| "held_out_episode_count": 14, |
| "metrics": { |
| "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 |
| } |
| } |
| }, |
| "deltas_v6_minus_v5": { |
| "json_validity_rate": -0.0009920634920634888, |
| "action_macro_f1": 0.0005933613618821745, |
| "subtask_accuracy": -0.007462686567164178, |
| "transition_accuracy": -0.0009920634920634888, |
| "next_action_accuracy": -0.010565240359218173, |
| "contact_accuracy": 0.03125, |
| "object_micro_f1": -0.009647761529319436 |
| }, |
| "wins_for_v6": [ |
| "action_macro_f1", |
| "contact_accuracy" |
| ], |
| "wins_for_v5": [ |
| "json_validity_rate", |
| "subtask_accuracy", |
| "transition_accuracy", |
| "next_action_accuracy", |
| "object_micro_f1" |
| ], |
| "interpretation": "v6 is the newest verified Qwen LoRA branch and is better for action macro-F1 and contact accuracy, but v5 remains the safer pinned release row for JSON perfection, subtask/next-action accuracy, transition accuracy, and object micro-F1." |
| } |
|
|
