Update final Qwen artifact docs

ARTIFACT_GUIDE.md

@@ -108,9 +108,10 @@ research project.
 | --- | --- |
 | [`results/omni_finetune/DATA_ACCESS_STATUS.md`](results/omni_finetune/DATA_ACCESS_STATUS.md) | Summarizes the data-readiness checks required before a held-out Qwen3-Omni pilot can report metrics. |
 | [`results/omni_finetune/MULTI_EPISODE_ACCESS_STATUS.md`](results/omni_finetune/MULTI_EPISODE_ACCESS_STATUS.md) | Documents the public multi-episode access path, selected 128-episode pilot plan, and data requirements. |
-| [`docs/data/omni_finetune_verified_result.json`](docs/data/omni_finetune_verified_result.json) | Compact verified summary for the first selected-episode Qwen3-Omni diagnostic pilot, including split counts, held-out metrics, and the quality-target caveat. |
+| [`docs/data/omni_finetune_verified_result.json`](docs/data/omni_finetune_verified_result.json) | Compact verified summary for the final selected-episode Qwen3-Omni diagnostic result, including split counts, held-out metrics, quality-target status, and adapter repo. |
 | [`results/omni_finetune/verified_public/`](results/omni_finetune/verified_public/) | Public-safe verified held-out result packages. These include metrics, predictions, reports, manifests, training metadata, validation summaries, and audit files, but not raw data or weights. |
-| [`results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_eval/analysis/ERROR_ANALYSIS.md`](results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_eval/analysis/ERROR_ANALYSIS.md) | Derived held-out error analysis by episode, action family, train-seen status, required-modality state, and object category for the validation-aware Qwen3-Omni diagnostic pilot. |
+| [`results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora_eval_test_full/`](results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora_eval_test_full/) | Final verified Qwen3-Omni public package with 448 held-out predictions, 99.78% JSON validity, metrics, reports, training metadata, validation summaries, and package audit. |
+| [`https://huggingface.co/cy0307/ropedia-qwen3-omni-lora-128ep`](https://huggingface.co/cy0307/ropedia-qwen3-omni-lora-128ep) | Public LoRA adapter weight repository for the final 128-episode Qwen3-Omni diagnostic run; raw Xperience-10M data and base Qwen weights remain excluded. |
 | [`results/omni_finetune/multi_episode_128_task_baselines/BASELINE_ALIGNMENT_REPORT.md`](results/omni_finetune/multi_episode_128_task_baselines/BASELINE_ALIGNMENT_REPORT.md) | Same-split 128-episode simple and neural metadata baselines for the 12 task ids, aligned to the 96/16/16 Qwen3-Omni split and explicit about raw-feature-only tasks. |
 | [`results/omni_finetune/multi_episode_128_task_baselines/summary_report.json`](results/omni_finetune/multi_episode_128_task_baselines/summary_report.json) | Machine-readable split counts, run configuration, simple metrics, neural metrics, and unsupported raw-feature markers for the aligned 128-episode baseline suite. |
 | [`scripts/omni/run_128_task_baselines.py`](scripts/omni/run_128_task_baselines.py) | Runner for the aligned 128-episode metadata/text baselines; it consumes the derived Qwen JSONL export locally but does not publish raw data, Qwen weights, or LoRA weights. |

EVALUATION_PROTOCOL.md

@@ -45,20 +45,20 @@ are not foundation models.
 
 ## Task Contracts
 
-| Task | Family | Unit | Input -> target | Primary metric | Minimal | Neural |
-| --- | --- | --- | --- | --- | ---: | ---: |
-| timeline_action | supervised classification | single window | current 20-frame all-feature window -> current action label | macro_f1 (higher better) | 0.0500 | 0.0148 |
-| timeline_subtask | supervised classification | single window | current 20-frame all-feature window -> current subtask label | macro_f1 (higher better) | 0.0506 | 0.0281 |
-| transition_detection | temporal diagnostic | single window | current 20-frame all-feature window -> action boundary versus steady | macro_f1 (higher better) | 0.6118 | 0.5862 |
-| next_action | short-horizon prediction | single window | current 20-frame all-feature window at time t -> action label at t + 20 frames | macro_f1 (higher better) | 0.0593 | 0.0419 |
-| hand_trajectory_forecast | trajectory regression | single window | current all-feature window -> future left/right hand 3D joints for 10 frames | mpjpe (lower better) | 0.8647 | 0.1079 |
-| contact_prediction | binary classification | single window | non-contact and non-caption feature blocks -> any body contact | macro_f1 (higher better) | 1.0000 | 1.0000 |
-| object_relevance | multi-label classification | single window | non-caption feature blocks -> current relevant object set | micro_f1 (higher better) | 0.1803 | 0.1679 |
-| caption_grounding | retrieval | caption query | caption object/interaction query plus candidate sensor windows -> matching time window | mrr (higher better) | 0.0160 | 0.0168 |
-| cross_modal_retrieval | retrieval | sensor query | motion, IMU, and camera query features -> matching depth/video window | top5_accuracy (higher better) | 0.3678 | 0.1983 |
-| modality_reconstruction | cross-modal regression | single window | motion, IMU, and camera features -> depth/video feature vector | r2 (higher better) | -0.0153 | -0.0102 |
-| temporal_order | pairwise diagnostic | adjacent window pair | two adjacent windows -> correct versus reversed order | f1 (higher better) | 0.5400 | 0.8520 |
-| misalignment_detection | pairwise diagnostic | paired modality window | motion side plus visual/depth side -> aligned versus shifted by 8 windows | f1 (higher better) | 0.5052 | 0.7153 |
+| Task | Artifact id | Family | Unit | Input -> target | Primary metric | Minimal | Neural |
+| --- | --- | --- | --- | --- | --- | ---: | ---: |
+| Action Recognition | `timeline_action` | supervised classification | single window | current 20-frame all-feature window -> current action label | macro_f1 (higher better) | 0.0500 | 0.0148 |
+| Procedure Step Recognition | `timeline_subtask` | supervised classification | single window | current 20-frame all-feature window -> current subtask label | macro_f1 (higher better) | 0.0506 | 0.0281 |
+| Action Boundary Detection | `transition_detection` | temporal diagnostic | single window | current 20-frame all-feature window -> action boundary versus steady | macro_f1 (higher better) | 0.6118 | 0.5862 |
+| Next-Action Prediction | `next_action` | short-horizon prediction | single window | current 20-frame all-feature window at time t -> action label at t + 20 frames | macro_f1 (higher better) | 0.0593 | 0.0419 |
+| Hand Trajectory Forecasting | `hand_trajectory_forecast` | trajectory regression | single window | current all-feature window -> future left/right hand 3D joints for 10 frames | mpjpe (lower better) | 0.8647 | 0.1079 |
+| Contact State Prediction | `contact_prediction` | binary classification | single window | non-contact and non-caption feature blocks -> any body contact | macro_f1 (higher better) | 1.0000 | 1.0000 |
+| Object Relevance Prediction | `object_relevance` | multi-label classification | single window | non-caption feature blocks -> current relevant object set | micro_f1 (higher better) | 0.1803 | 0.1679 |
+| Language Grounding | `caption_grounding` | retrieval | caption query | caption object/interaction query plus candidate sensor windows -> matching time window | mrr (higher better) | 0.0160 | 0.0168 |
+| Cross-Modal Retrieval | `cross_modal_retrieval` | retrieval | sensor query | motion, IMU, and camera query features -> matching depth/video window | top5_accuracy (higher better) | 0.3678 | 0.1983 |
+| Cross-Modal Reconstruction | `modality_reconstruction` | cross-modal regression | single window | motion, IMU, and camera features -> depth/video feature vector | r2 (higher better) | -0.0153 | -0.0102 |
+| Temporal Order Verification | `temporal_order` | pairwise diagnostic | adjacent window pair | two adjacent windows -> correct versus reversed order | f1 (higher better) | 0.5400 | 0.8520 |
+| Multimodal Synchronization Detection | `misalignment_detection` | pairwise diagnostic | paired modality window | motion side plus visual/depth side -> aligned versus shifted by 8 windows | f1 (higher better) | 0.5052 | 0.7153 |
 
 ## Leakage Controls
 
@@ -72,7 +72,7 @@ are not foundation models.
 
 - Cross-episode generalization for Qwen3-Omni has a first verified diagnostic pilot, but strong model quality is not yet shown.
 - Feature-vector reconstruction is separate from pixel depth, mesh, NeRF, or Gaussian reconstruction.
-- The verified validation-aware Qwen3-Omni diagnostic pilot has weak held-out metrics and needs structured-output and task-quality improvements before larger model-quality claims.
+- The final verified Qwen3-Omni diagnostic result meets the strict-JSON target, but action/subtask held-out quality remains weak and needs error analysis before larger model-quality claims.
 - Full audio-visual representation learning still needs multi-episode training; the current report includes single-episode audio/no-audio ablations.
 
 ## Scale-Up Gate
@@ -86,7 +86,7 @@ claiming improved held-out model quality:
 - manifest, training metadata, progress logs, metrics, predictions, and run report
 - held-out evaluation on test episodes rather than train windows
 
-Current status: verified diagnostic pilot; quality target not met. Read
+Current status: verified diagnostic result; strict-JSON quality target met, action/subtask quality still weak. Read
 `docs/data/omni_finetune_verified_result.json` before interpreting any
 Qwen3-Omni metric.
 

PROJECT_BRIEF.md

@@ -21,7 +21,7 @@ results, and see what remains before multi-episode model-quality claims.
 | Data understanding | `feature_manifest.json`, `available_modalities.json`, modality atlas, episode-window HF viewer |
 | Task design | 12 task contracts, task cards, case-study walkthroughs, and four research-direction extension probes |
 | Evaluation rigor | chronological split, per-task metrics, predictions, confusion matrices, leakage notes, and generated takeaways |
-| Scale-up planning | 128-episode selection/relay plan, Qwen3-Omni path, Cosmos 3 branch, and policy-model candidates after action-space conversion |
+| Scale-up planning | Final verified 96/16/16 Qwen3-Omni diagnostic result, same-split 128-episode baseline alignment, Cosmos3-Nano compatibility branch, and policy-model candidates after action-space conversion |
 
 ## What Exists Now
 
@@ -32,7 +32,7 @@ results, and see what remains before multi-episode model-quality claims.
 | Task suite | 12 embodied-AI task contracts with inputs, targets, metrics, predictions, and case-study walkthroughs |
 | Models | Minimal linear/ridge/logistic baselines plus compact PyTorch MLP heads for the same 12 tasks |
 | Research map | Four Ropedia research directions with direct, proxy, diagnostic, and extension-task coverage |
-| Scale-up path | Qwen3-Omni LoRA code path prepared; the gated Xperience-10M dataset is available for a selected 128-episode pilot |
+| Scale-up path | A selected 96/16/16 Qwen3-Omni LoRA final diagnostic result is verified; strict-JSON validity meets target, while weak action/subtask metrics guide the next error-analysis pass |
 
 ## How To Read It
 
@@ -42,15 +42,17 @@ results, and see what remains before multi-episode model-quality claims.
 3. Open `EVALUATION_PROTOCOL.md` before comparing task scores.
 4. Use `RESEARCH_TAKEAWAYS.md` for the current metric interpretation.
 5. Inspect `results/episode_task_suite/feature_manifest.json` to understand one model input.
-6. Use `results/omni_finetune/DATA_ACCESS_STATUS.md` for the multi-episode data status.
+6. Use `docs/data/omni_finetune_verified_result.json` for the current multi-episode Qwen3-Omni pilot result.
 
 ## What This Enables
 
 The public sample is enough to build and verify task definitions, feature
 contracts, metrics, visualization, and baseline code. It is not enough to
-measure final model quality for a general embodied-AI model. The next research
-stage is to run the same contracts on held-out episodes, then fine-tune and
-evaluate an omni-model with train/test separation at the episode level.
+measure final model quality for a general embodied-AI model. The first
+multi-episode Qwen3-Omni diagnostic pilot now verifies the held-out training
+loop with validation loss recorded; the next research stage is to improve
+JSON-format reliability and error analysis before larger robustness or
+alternative backbone claims.
 
 ## Best Entry Points
 

PROJECT_STATUS.md

@@ -21,7 +21,7 @@ scale-up readiness; it is not presented as final full-dataset model quality.
 | Neural heads | Verified | `scripts/neural_task_models.py`, `results/episode_task_suite/neural_mlp/` | Each task also has a compact PyTorch MLP run over the same feature tensor and chronological split. |
 | Audio contribution study | Verified | `scripts/audio_ablation_and_raw_upgrade.py`, `results/audio_ablation/`, `docs/data/audio_ablation_summary.json` | Audio variants are compared across all 12 task contracts; audio improves the primary metric on 6 of 12 tasks, and a 588-d audio-window representation improves over the baseline audio variant on 6 of 12 tasks. |
 | Research takeaways | Verified | `RESEARCH_TAKEAWAYS.md`, `docs/data/research_takeaways.json`, `scripts/build_research_takeaways.py` | The main result interpretation is generated from committed metrics: chronological class shift, neural gains on dynamics/order/alignment, open retrieval/reconstruction problems, and the need for held-out episodes. |
-| Research roadmap | Current | `RESEARCH_ROADMAP.md`, `docs/data/research_roadmap.json` | The roadmap connects public-sample task development to the verified validation-aware Qwen3-Omni diagnostic baseline, structured-output improvement pass, robustness runs, world/policy branches, and the future Xperience-native pretraining goal. |
+| Research roadmap | Current | `RESEARCH_ROADMAP.md`, `docs/data/research_roadmap.json` | The roadmap connects public-sample task development to the final verified Qwen3-Omni diagnostic result, same-split baseline alignment, action/subtask error analysis, robustness runs, world/policy branches, and the future Xperience-native pretraining goal. |
 | Foundation-model plan | Current | `FOUNDATION_MODEL_PLAN.md`, `docs/data/foundation_model_plan.json` | Qwen3-Omni remains the first trainable held-out LoRA baseline; Cosmos 3 is added as the first world-model/action-generation branch; OpenVLA/openpi/GR00T are policy candidates after action targets are explicit. |
 | Omni model extension contract | Current | `OMNI_MODEL_EXTENSION_CONTRACT.md`, `configs/omni_backbones/`, `scripts/omni/backbone_registry.py`, `scripts/omni/smoke_test_backbone_packaging.py` | Future model branches must keep the same episode split discipline, held-out metrics, validation gate, public-safe package contract, and explicit forbidden-artifact policy before reporting results. |
 | Xperience Embodied Foundation Model | Future goal | `XPERIENCE_EMBODIED_FOUNDATION_MODEL_PRETRAINING.md` | A future full-corpus pretraining plan describes target modules, objectives, staged scale-up, hardware ranges, and evaluation for a domain-specific embodied foundation model. |
@@ -31,7 +31,7 @@ scale-up readiness; it is not presented as final full-dataset model quality.
 | Public package policy | Verified | `DATA_NOTICE.md`, `REPRODUCIBILITY.md` | Raw Xperience-10M data, private gated files, large archives, credentials, and full Qwen weights are not redistributed. |
 | Reproducibility | Verified for the public sample | `REPRODUCIBILITY.md`, `docs/data/reproducibility_matrix.json`, `notes/reproducibility_audit.md` | The public sample workflow has explicit commands, expected outputs, and exact-match reproduction evidence. |
 | 128-episode aligned baselines | Verified companion result | `results/omni_finetune/multi_episode_128_task_baselines/BASELINE_ALIGNMENT_REPORT.md`, `results/omni_finetune/multi_episode_128_task_baselines/summary_report.json`, `scripts/omni/run_128_task_baselines.py` | The earlier simple and neural baseline framing is aligned to the same selected 96/16/16 episode split used by the Qwen3-Omni pilot. JSON-supported tasks have metadata/text simple and neural MLP metrics; raw-feature-only tasks are explicitly marked unsupported until 128-run sensor feature blocks are available. |
-| Qwen3-Omni fine-tuning | Verified validation-aware diagnostic held-out pilot; quality target not met | `docs/data/omni_finetune_verified_result.json`, `results/omni_finetune/verified_public/`, `results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_eval/analysis/`, `scripts/omni/package_verified_omni_result.py`, `scripts/omni/audit_verified_omni_package.py`, `scripts/omni/analyze_qwen3_omni_errors.py` | The selected 96/16/16 episode split produced a validation-aware public-safe held-out package with 3,808 exported windows, 512 validation windows, 448 test predictions, and derived error-analysis tables by episode, action family, train-seen status, required-modality state, and object category. JSON validity is 87.50%, below the 98% target, so the result is a diagnostic baseline and the next pass should focus on structured-output improvements. |
+| Qwen3-Omni fine-tuning | Final verified diagnostic held-out result; JSON target met | `docs/data/omni_finetune_verified_result.json`, `results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora_eval_test_full/`, `scripts/omni/package_verified_omni_result.py`, `scripts/omni/audit_verified_omni_package.py`, `scripts/omni/analyze_qwen3_omni_errors.py` | The selected 96/16/16 episode split produced a final public-safe held-out package with 3,808 exported windows, 512 validation windows, 448 test predictions, two training epochs, and validation/audit summaries. JSON validity is 99.78%, meeting the 98% target; transition accuracy is 97.10%, contact accuracy is 71.88%, object micro-F1 is 30.16%, and action/subtask metrics remain weak, so it is still a diagnostic baseline rather than a strong model-quality claim. |
 | Raw Xperience-10M redistribution | Not included | `DATA_NOTICE.md`, `docs/data/publication_audit.json` | Raw MP4, HDF5, RRD files, private gated data, and full Qwen weights are intentionally excluded. |
 
 ## Fast Research Route
@@ -69,9 +69,9 @@ scale-up readiness; it is not presented as final full-dataset model quality.
   current results use one public sample episode.
 - Public-facing fine-tuning results should come from the verified result
   package, not from live process logs or setup-only artifacts.
-- The first Qwen3-Omni held-out package verifies the pipeline, not strong model
-  quality: JSON validity is 87.50%, action macro-F1 is 0.0027, and subtask
-  accuracy is 0.0067.
+- The final Qwen3-Omni held-out package verifies the pipeline and meets the
+  strict-JSON target, but not strong action/subtask model quality: JSON validity
+  is 99.78%, action macro-F1 is 0.0024, and subtask accuracy is 0.0022.
 - The 128-episode aligned simple/NN baselines use metadata/text features from
   the derived Qwen JSONL export; they align the split and task ids but do not
   replace raw-modality baselines for trajectory, retrieval, reconstruction, or

RESEARCH_ROADMAP.md

@@ -11,11 +11,11 @@ should exist before the stage is treated as complete.
 | --- | --- | --- | --- | --- |
 | Public-Sample Task Lab | Implemented | One public Xperience-10M sample episode is available. | 1,161 aligned windows, 12 task contracts, minimal heads, neural MLP heads, modality atlas, task walkthroughs, and derived figures. | `PROJECT_STATUS.md`, `EVALUATION_PROTOCOL.md`, `RESEARCH_TAKEAWAYS.md`, `docs/data/summary_metrics.json`, `results/episode_task_suite/summary_report.json` |
 | Multi-Episode Data Preparation | Implemented for first selected pilot | Gated dataset availability and enough storage for selected episodes. | 128 selected episodes, episode manifest, missing-view manifest, held-out episode split, and source-discovery report. | `results/omni_finetune/DATA_ACCESS_STATUS.md`, `results/omni_finetune/MULTI_EPISODE_ACCESS_STATUS.md`, `results/omni_finetune/xperience10m_128_episode_selection.json` |
-| Qwen3-Omni LoRA Validation-Aware Diagnostic Pilot | Verified baseline | Selected episodes prepared locally with no train/test episode leakage. | Dataset JSONL/media manifests, LoRA adapter checkpoint, progress logs, validation monitoring, held-out predictions, metrics, confusion matrices, and run report. | `docs/data/omni_finetune_verified_result.json`, `results/omni_finetune/verified_public/`, `metrics.json`, `predictions.jsonl`, `RUN_REPORT.md` |
+| Qwen3-Omni LoRA Final Diagnostic Result | Verified baseline | Selected episodes prepared locally with no train/test episode leakage. | Dataset JSONL/media manifests, LoRA adapter checkpoint, progress logs, validation monitoring, held-out predictions, metrics, confusion matrices, run report, and public LoRA adapter repo. | `docs/data/omni_finetune_verified_result.json`, `results/omni_finetune/verified_public/`, `metrics.json`, `predictions.jsonl`, `RUN_REPORT.md`, `https://huggingface.co/cy0307/ropedia-qwen3-omni-lora-128ep` |
 | 128-Episode Same-Split Simple/NN Baselines | Verified companion result | Derived Qwen JSONL export for the selected 96/16/16 split. | Same 12 task ids, simple metadata/text baselines, neural MLP baselines where JSON labels support them, and explicit unsupported markers for tasks that still require raw 128 feature blocks. | `results/omni_finetune/multi_episode_128_task_baselines/BASELINE_ALIGNMENT_REPORT.md`, `summary_report.json`, `scripts/omni/run_128_task_baselines.py` |
-| Structured-Output And Error-Analysis Pass | Active next step | The validation-aware diagnostic package exists and shows weak held-out quality. | Same 96/16/16 split, stricter JSON decoding or target formatting, action/subtask error analysis, held-out test evaluation, and comparison to the verified validation-aware baseline. | Updated quality-target report, error-analysis tables, held-out metrics, and verified public package. |
+| Action/Subtask Error-Analysis Pass | Active next step | The final diagnostic package meets strict JSON validity but has weak action/subtask held-out quality. | Same 96/16/16 split, action/subtask confusion analysis, unseen-label analysis, object/action family breakdowns, and comparison to the final verified Qwen baseline. | Updated error-analysis tables, held-out metrics by failure type, and verified public package. |
 | Foundation-Model Selection Matrix | Current | The selected pilot episodes are prepared, or a 3-8 episode dry run is available for preprocessing checks. | Backbone registry, Cosmos 3 world-model branch plan, Qwen3-Omni baseline plan, OpenVLA/openpi/GR00T policy candidates, and model-specific evaluation additions. | `FOUNDATION_MODEL_PLAN.md`, `docs/data/foundation_model_plan.json`, `research_roadmap_interactive.json` |
-| 64-128 Episode Robustness Run | Planned | The validation-aware selected-episode pilot trains and evaluates cleanly. | Split-by-session metrics, modality ablations, calibration/object/language error analysis, and sensitivity to missing views. | Held-out metrics by session, task, and modality; ablation tables; qualitative error analysis. |
+| 64-128 Episode Robustness Run | Planned | The final selected-episode Qwen diagnostic run trains and evaluates cleanly. | Split-by-session metrics, modality ablations, calibration/object/language error analysis, and sensitivity to missing views. | Held-out metrics by session, task, and modality; ablation tables; qualitative error analysis. |
 | Cosmos 3 and Policy-Model Extensions | Planned | Enough multi-episode data, compute budget, and model-specific action/world-state targets. | Cosmos 3 future-window or action-conditioned world-model probes, OpenVLA/openpi/GR00T action-policy baselines, modality-conditioning checks, affordance tasks, and synthetic-data usefulness tests. | Task-specific held-out evaluations, qualitative inspection, and updated model cards. |
 | Xperience Embodied Foundation Model Pretraining | Future | Full-corpus access, PB-scale storage path, multi-node compute, and positive scaling evidence from smaller runs. | Xperience-native temporal multimodal model, full-corpus manifests, pretraining shards, scaling curves, held-out evaluations, and model card. | Pretraining metadata, checkpoint inventory, held-out metrics, scaling report, and data-boundary report. |
 
@@ -23,8 +23,8 @@ should exist before the stage is treated as complete.
 
 The useful next decision is model-quality improvement plus backbone fit: keep
 the public-sample task suite as the development harness, use the verified
-Qwen3-Omni validation-aware diagnostic pilot as the first cross-episode
-baseline, then improve format reliability and task quality before claiming
+Qwen3-Omni final diagnostic result as the first cross-episode
+baseline, then improve action/subtask quality before claiming
 model quality. The earlier simple and neural baseline framing is now aligned to
 the same 96/16/16 split through metadata/text baselines for JSON-supported task
 ids; raw-feature-only tasks remain marked as needing the 128-run sensor feature
@@ -93,11 +93,12 @@ Evidence to inspect:
 ### 3. Qwen3-Omni LoRA Pilot
 
 This stage uses Qwen3-Omni as the multimodal backbone and trains lightweight
-LoRA adapters. The first held-out diagnostic package now exists. It proves the
-export, training, evaluation, validation, and public-safe packaging loop, but
-the metrics are weak: JSON validity is 87.50%, action macro-F1 is 0.0027, and
-subtask accuracy is 0.0067. Treat it as a baseline and error-analysis starting
-point.
+LoRA adapters. The final held-out diagnostic package now exists. It proves the
+export, training, evaluation, validation, public-safe packaging, and adapter
+publication loop. JSON validity is 99.78%, transition accuracy is 97.10%, and
+contact accuracy is 71.88%, but action macro-F1 is 0.0024 and subtask accuracy
+is 0.0022. Treat it as a baseline and error-analysis starting point, not as a
+strong action/subtask model.
 
 Expected outputs:
 

RESEARCH_TAKEAWAYS.md

@@ -97,18 +97,18 @@ Current scope: This is a single-episode ablation over fixed ridge heads. It vali
 
 ### The next scientific unit is held-out episodes, not more adjacent windows
 
-The selected Qwen3-Omni path now has a verified validation-aware held-out diagnostic pilot. It proves the cross-episode train/validation/eval loop, but the weak metrics show that structured-output reliability and task-quality error analysis are the next modeling problems.
+The selected Qwen3-Omni path now has a verified two-epoch held-out diagnostic result. It proves the cross-episode train/validation/eval loop and meets the strict-JSON target, while weak action/subtask metrics remain the next modeling problem.
 
 | Metric | Value |
 | --- | ---: |
 | `selected_episodes` | 128 |
-| `held_out_test_windows` | 448 |
-| `json_validity_rate` | 0.8750 |
-| `action_macro_f1` | 0.0027 |
+| `held_out_test_windows` | n/a |
+| `json_validity_rate` | n/a |
+| `action_macro_f1` | n/a |
 
 Source: `docs/data/omni_finetune_verified_result.json`.
 
-Current scope: The selected-episode Qwen3-Omni validation-aware diagnostic pilot is verified, but held-out quality is still weak and JSON validity remains below the 98% target.
+Current scope: The selected-episode Qwen3-Omni diagnostic pilot is verified on the 96/16/16 split and now meets the 98% target for JSON validity; action/subtask quality remains weak, so current results are diagnostic baselines, not strong model-quality claims.
 
 ## How To Read These Results