| { |
| "current_public_matrix_row": "qwen3_omni_v6_lora", |
| "generated_at_utc": "2026-06-21T11:47:45+00:00", |
| "interpretation_rule": "Do not confuse the Qwen run versions with the project evidence lines. The project evidence lines are one public sample episode and selected 128-episode artifacts. Qwen v1-v6 are only the Qwen3-Omni run lineage inside the selected-128 line. The 20-task matrix uses Qwen3-Omni v6 LoRA; v5 remains the pinned prior release; v1-v4 are lineage and ablation evidence.", |
| "pinned_prior_release": "v5", |
| "related_engineering_artifacts": [ |
| { |
| "name": "Full-parameter gates", |
| "path": "results/omni_finetune/QWEN3_FULL_PARAMETER_GATES_20260609.md", |
| "role": "Feasibility and short-train gates; not a public 20-task matrix method row." |
| }, |
| { |
| "name": "Alternate fullsplit v6 package", |
| "path": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu_lora_fsdp_full_train_noval_tail_logits_fullstatesave_v6_eval_test_full", |
| "role": "Verified alternate no-validation/fullsplit artifact retained for audit, not the current matrix row." |
| } |
| ], |
| "runs": [ |
| { |
| "change_from_previous": "First verified Qwen3-Omni selected-128 LoRA run.", |
| "dataset_contract": "xperience10m_episode_json_qa_v1", |
| "dataset_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605", |
| "eval_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_eval", |
| "eval_samples": 448, |
| "metrics": { |
| "action_macro_f1": 0.0026621494447581404, |
| "contact_accuracy": 0.6450892857142857, |
| "json_validity_rate": 0.875, |
| "next_action_accuracy": 0.024553571428571428, |
| "object_micro_f1": 0.22299431459254582, |
| "subtask_accuracy": 0.006696428571428571, |
| "transition_accuracy": 0.8504464285714286 |
| }, |
| "package": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_eval", |
| "package_path": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_eval", |
| "public_matrix_role": "superseded lineage evidence, not the current 20-task Qwen row", |
| "purpose": "Prove that the selected-128 split, LoRA training, held-out eval, validation, and public packaging loop works end to end.", |
| "reader_use": "Use only as lineage evidence for the first working pipeline.", |
| "role": "First verified 96/16/16 selected-episode Qwen3-Omni LoRA package; establishes dataset, training, eval, and packaging plumbing.", |
| "status": "verified", |
| "title": "Selected-128 validation-aware LoRA baseline", |
| "train_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_lora", |
| "version": "v1" |
| }, |
| { |
| "change_from_previous": "Reused the selected-128 split with a stricter structured-JSON answer contract and full 8-GPU LoRA training.", |
| "dataset_contract": "xperience10m_episode_json_qa_v1", |
| "dataset_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605", |
| "eval_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora_eval_test_full", |
| "eval_samples": 448, |
| "metrics": { |
| "action_macro_f1": 0.0024331644885523347, |
| "contact_accuracy": 0.71875, |
| "json_validity_rate": 0.9977678571428571, |
| "next_action_accuracy": 0.029017857142857144, |
| "object_micro_f1": 0.30160427807486634, |
| "subtask_accuracy": 0.002232142857142857, |
| "transition_accuracy": 0.9709821428571429 |
| }, |
| "package": "xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora_eval_test_full", |
| "package_path": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora_eval_test_full", |
| "public_matrix_role": "superseded lineage evidence, not the current 20-task Qwen row", |
| "purpose": "Make the answer format schema-checked and reduce invalid JSON before expanding scale.", |
| "reader_use": "Use as evidence that schema-constrained evaluation improved validity and contact accuracy over v1.", |
| "role": "Reuses the selected-128 split with a stricter structured JSON answer contract and full 8-GPU LoRA training.", |
| "status": "verified", |
| "title": "Structured-JSON reuse full-8-GPU LoRA", |
| "train_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora", |
| "version": "v2" |
| }, |
| { |
| "change_from_previous": "Evaluated the v2 adapter with stricter labels and prompts; no new adapter training.", |
| "dataset_contract": "xperience10m_episode_json_qa_v1", |
| "dataset_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605", |
| "eval_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v3_strict_label_prompt_reuse_lora_eval_test_full", |
| "eval_samples": 448, |
| "metrics": { |
| "action_macro_f1": 0.0021983997167007384, |
| "contact_accuracy": 0.7209821428571429, |
| "json_validity_rate": 1.0, |
| "next_action_accuracy": 0.03125, |
| "object_micro_f1": 0.30688228657389993, |
| "subtask_accuracy": 0.002232142857142857, |
| "transition_accuracy": 0.9732142857142857 |
| }, |
| "package": "xperience10m_qwen3_omni_128ep_structured_json_v3_strict_label_prompt_reuse_lora_eval_test_full", |
| "package_path": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v3_strict_label_prompt_reuse_lora_eval_test_full", |
| "public_matrix_role": "superseded prompt/eval lineage evidence", |
| "purpose": "Separate prompt/eval formatting effects from adapter-training effects.", |
| "reader_use": "Use as prompt/eval ablation evidence, not as a separate trained model.", |
| "role": "Strict-label prompt/eval pass over the v2 adapter; improves JSON validity without introducing a new adapter training run.", |
| "status": "verified", |
| "title": "Strict-label prompt evaluation", |
| "train_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora", |
| "version": "v3" |
| }, |
| { |
| "change_from_previous": "Trained a new four-epoch full-8-GPU LoRA adapter on the structured-JSON setup.", |
| "dataset_contract": "xperience10m_episode_json_qa_v1", |
| "dataset_run_id": "xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605", |
| "eval_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora_eval_test_full", |
| "eval_samples": 448, |
| "metrics": { |
| "action_macro_f1": 0.0018678269676001454, |
| "contact_accuracy": 0.7299107142857143, |
| "json_validity_rate": 1.0, |
| "next_action_accuracy": 0.033482142857142856, |
| "object_micro_f1": 0.31099781500364165, |
| "subtask_accuracy": 0.0, |
| "transition_accuracy": 0.9732142857142857 |
| }, |
| "package": "xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora_eval_test_full", |
| "package_path": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora_eval_test_full", |
| "public_matrix_role": "superseded lineage evidence, not the current 20-task Qwen row", |
| "purpose": "Test whether longer structured-JSON LoRA training improves the same selected split.", |
| "reader_use": "Use as overfit and metric-tradeoff evidence before the multiscale export.", |
| "role": "Four-epoch full-8-GPU LoRA run on the same selected split; useful for overfit/metric tradeoff analysis.", |
| "status": "verified", |
| "title": "Four-epoch structured-JSON LoRA", |
| "train_run_id": "xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora", |
| "version": "v4" |
| }, |
| { |
| "change_from_previous": "Introduced the multiscale cap96 export and larger held-out evaluation surface.", |
| "dataset_contract": "xperience10m_episode_json_qa_v1", |
| "dataset_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", |
| "eval_samples": 4032, |
| "metrics": { |
| "action_macro_f1": 0.002289711036077459, |
| "contact_accuracy": 0.7864583333333334, |
| "json_validity_rate": 1.0, |
| "next_action_accuracy": 0.053618594823032224, |
| "object_micro_f1": 0.31614599936244814, |
| "subtask_accuracy": 0.011194029850746268, |
| "transition_accuracy": 0.9908234126984127 |
| }, |
| "package": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora_eval_test_full", |
| "package_path": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora_eval_test_full", |
| "public_matrix_role": "pinned prior release row and comparison baseline", |
| "purpose": "Move from the 448-sample compact eval to a denser multiscale 4,032-sample held-out eval.", |
| "reader_use": "Use as the pinned prior release; it remains stronger on JSON validity, subtask, next-action, object, and transition metrics.", |
| "role": "Dense/multiscale selected-128 run with 4,032 held-out predictions; kept as the pinned prior release because several metrics remain stronger than v6.", |
| "status": "verified", |
| "title": "Multiscale cap96 LoRA", |
| "train_run_id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora", |
| "version": "v5" |
| }, |
| { |
| "change_from_previous": "Kept the multiscale setup, changed LoRA rank/lr to rank64/lr5e-5, and added verified task-specific probes for full 20-task coverage.", |
| "dataset_contract": "xperience10m_episode_json_qa_v1", |
| "dataset_run_id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora", |
| "eval_run_id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora_eval_test_full", |
| "eval_samples": 4032, |
| "metrics": { |
| "action_macro_f1": 0.0028830723979596335, |
| "contact_accuracy": 0.8177083333333334, |
| "json_validity_rate": 0.9990079365079365, |
| "next_action_accuracy": 0.04305335446381405, |
| "object_micro_f1": 0.3064982378331287, |
| "subtask_accuracy": 0.0037313432835820895, |
| "transition_accuracy": 0.9898313492063492 |
| }, |
| "package": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora_eval_test_full", |
| "package_path": "results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora_eval_test_full", |
| "public_matrix_role": "current public 20-task Qwen3-Omni v6 LoRA row", |
| "purpose": "Promote the current public Qwen3-Omni 20-task row with multiscale LoRA plus task-specific probes.", |
| "reader_use": "Use as the current public 20-task Qwen row; it improves action macro-F1 and contact accuracy while v5 remains the prior comparator.", |
| "role": "Current verified Qwen3-Omni row: rank64/lr5e-5 multiscale LoRA plus task-specific probe artifacts used for the 20/20 Qwen matrix coverage.", |
| "status": "verified", |
| "title": "Rank64 lr5e-5 multiscale LoRA", |
| "train_run_id": "xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora", |
| "version": "v6" |
| } |
| ], |
| "scope": "Verified public-safe Qwen3-Omni LoRA/eval packages over the selected Xperience-10M 128-episode surface.", |
| "status": "pass", |
| "title": "Qwen3-Omni v1-v6 Run Lineage" |
| } |
|
|
