results/omni_finetune/model_output_task_probes_20260616/RUN_REPORT.md

Existing Model-Output Task Probes

Generated: 2026-06-18T22:52:18+00:00

This package scores only task targets already present in verified held-out prediction JSON. It does not run new inference and does not infer targets that are absent from a model branch.

Method	ID	Status	Scored tasks	Task 13 macro-F1	Task 14 macro-F1	Task 16 macro-F1	Task 17 micro-F1	Task 20 MAE	Task 8 IoU	Evidence
Qwen3-Omni v6 LoRA	qwen3_omni_v6_lora	scored	action_object_relation	n/a	n/a	0.000222	n/a	n/a	n/a	results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora_eval_test_full/eval/predictions.jsonl
Cosmos3-Super Reasoner	cosmos3_super_reasoner	scored	action_object_relation, caption_grounding, long_horizon_next_action, time_to_transition	0.008808	n/a	0.000000	n/a	52.946	0.306399	results/omni_finetune/verified_public/xperience10m_cosmos3_super_reasoner_128ep_test_full_20260607/eval/predictions.jsonl
Cosmos3-Nano Future Window	cosmos3_nano_future_window	scored	action_object_relation, long_horizon_next_action, modality_reconstruction, next_subtask_forecast, object_set_forecast, time_to_transition	0.002491	0.006615	0.002794	0.017820	33.810	n/a	results/omni_finetune/verified_public/xperience10m_cosmos3_nano_128ep_future_window_h5_compat_adapter_eval_test_full/eval/future_predictions.jsonl