Views
No views yet
1cd .../submit_materials
2
3DATASET_ROOT=/path/to/PortraitCraft \
4CUDA_VISIBLE_DEVICES=0,1,2 \
5WORKERS_PER_GPU=2 \
6bash run_reproduce.shDATASET_ROOT must contain track_1_test.json and the image shard folders
images_00, images_01, ..., images_10. You can override the image search
root with IMAGE_DIR=/path/to/image_shard_root.1RUN_QA_INFERENCE=0 RUN_FINE_LEVEL_INFERENCE=0 \
2DATASET_ROOT=/path/to/PortraitCraft \
3bash run_reproduce.shRUN_QA_INFERENCE=0, the script reuses the existing final QA file when it
is present. Set FORCE_QA_POSTPROCESS=1 to force QA postprocessing again.
If RUN_FINE_LEVEL_INFERENCE=0, it similarly reuses existing final criteria
and total-score files when present. Set FORCE_FINE_LEVEL_POSTPROCESS=1 to
force fine-level postprocessing again.track_1_test.jsontotal_score: outputs_fine_level/track_1_test_fine_level_tta_mean_then_calibrate_all13.jsoncriteria: outputs_fine_level/track_1_test_fine_level_tta_mean_then_affine_gamma1p27_all13.jsonanswer: outputs_qa/tta9/track_1_test_qa_answer_tta9_lowconf_logprob_m2_s1_second.jsonimage_path, total_score, criteria, question, options, answermodels/Qwen3.5-4B/: base model.lora/checkpoint-500/: QA LoRA.lora/fine_level_checkpoint-100/: fine-level LoRA used for criteria and total_score.code/merge_track1_submission.py: final merge step.outputs_debug: local reproduction and running results.