QWEN3_OMNI_RUN_LINEAGE.md

Qwen3-Omni v1-v6 Run Lineage

Generated: 2026-06-21T11:47:45+00:00.

Scope: Verified public-safe Qwen3-Omni LoRA/eval packages over the selected Xperience-10M 128-episode surface.

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.

Read the versions as an engineering audit trail, not as six separate benchmark rows. v1-v4 explain how the Qwen3-Omni pipeline was hardened, v5 is the pinned prior multiscale release, and v6 is the current 20-task Qwen3-Omni row.

Compact Lineage

Version	Run	Purpose	Change from previous	Eval samples	JSON validity	Action macro-F1	Contact acc.	Use now
v1	Selected-128 validation-aware LoRA baseline	Prove that the selected-128 split, LoRA training, held-out eval, validation, and public packaging loop works end to end.	First verified Qwen3-Omni selected-128 LoRA run.	448	0.8750	0.0027	0.6451	Use only as lineage evidence for the first working pipeline.
v2	Structured-JSON reuse full-8-GPU LoRA	Make the answer format schema-checked and reduce invalid JSON before expanding scale.	Reused the selected-128 split with a stricter structured-JSON answer contract and full 8-GPU LoRA training.	448	0.9978	0.0024	0.7188	Use as evidence that schema-constrained evaluation improved validity and contact accuracy over v1.
v3	Strict-label prompt evaluation	Separate prompt/eval formatting effects from adapter-training effects.	Evaluated the v2 adapter with stricter labels and prompts; no new adapter training.	448	1.0000	0.0022	0.7210	Use as prompt/eval ablation evidence, not as a separate trained model.
v4	Four-epoch structured-JSON LoRA	Test whether longer structured-JSON LoRA training improves the same selected split.	Trained a new four-epoch full-8-GPU LoRA adapter on the structured-JSON setup.	448	1.0000	0.0019	0.7299	Use as overfit and metric-tradeoff evidence before the multiscale export.
v5	Multiscale cap96 LoRA	Move from the 448-sample compact eval to a denser multiscale 4,032-sample held-out eval.	Introduced the multiscale cap96 export and larger held-out evaluation surface.	4032	1.0000	0.0023	0.7865	Use as the pinned prior release; it remains stronger on JSON validity, subtask, next-action, object, and transition metrics.
v6	Rank64 lr5e-5 multiscale LoRA	Promote the current public Qwen3-Omni 20-task row with multiscale LoRA plus task-specific probes.	Kept the multiscale setup, changed LoRA rank/lr to rank64/lr5e-5, and added verified task-specific probes for full 20-task coverage.	4032	0.9990	0.0029	0.8177	Use as the current public 20-task Qwen row; it improves action macro-F1 and contact accuracy while v5 remains the prior comparator.

Run IDs And Packages

Version	Train run	Eval run	Role	Package
v1	xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_lora	xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_eval	First verified 96/16/16 selected-episode Qwen3-Omni LoRA package; establishes dataset, training, eval, and packaging plumbing.	results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_96train_16val_16test_valmon_20260605_eval
v2	xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora	xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora_eval_test_full	Reuses the selected-128 split with a stricter structured JSON answer contract and full 8-GPU LoRA training.	results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora_eval_test_full
v3	xperience10m_qwen3_omni_128ep_structured_json_v2_reuse_full8gpu_lora	xperience10m_qwen3_omni_128ep_structured_json_v3_strict_label_prompt_reuse_lora_eval_test_full	Strict-label prompt/eval pass over the v2 adapter; improves JSON validity without introducing a new adapter training run.	results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v3_strict_label_prompt_reuse_lora_eval_test_full
v4	xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora	xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora_eval_test_full	Four-epoch full-8-GPU LoRA run on the same selected split; useful for overfit/metric tradeoff analysis.	results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_structured_json_v4_4epoch_full8gpu_lora_eval_test_full
v5	xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora	xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora_eval_test_full	Dense/multiscale selected-128 run with 4,032 held-out predictions; kept as the pinned prior release because several metrics remain stronger than v6.	results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora_eval_test_full
v6	xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora	xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora_eval_test_full	Current verified Qwen3-Omni row: rank64/lr5e-5 multiscale LoRA plus task-specific probe artifacts used for the 20/20 Qwen matrix coverage.	results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora_eval_test_full

Related Engineering Artifacts

Artifact	Path	Role
Full-parameter gates	results/omni_finetune/QWEN3_FULL_PARAMETER_GATES_20260609.md	Feasibility and short-train gates; not a public 20-task matrix method row.
Alternate fullsplit v6 package	results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu_lora_fsdp_full_train_noval_tail_logits_fullstatesave_v6_eval_test_full	Verified alternate no-validation/fullsplit artifact retained for audit, not the current matrix row.