PhyRubric-Train

📄 Paper: PhyRubric: Rubric-Grounded Diagnostic Evaluation of Physical Commonsense in Text-to-Video Generation

The training videos of PhyRubric, a rubric-grounded diagnostic framework for physical commonsense in text-to-video generation.

947 captions (disjoint from the 245 of PhyRubric-Bench) across the same five physical domains, rendered by four video generators: 3,788 clips, every one labelled item by item against its caption-specific rubric and scored for semantic adherence (SA) and physical commonsense (PC) on 1-5. PhyRubric-Judge is fine-tuned on these labels.

<generator>/train/<ID>_s42.mp4       3,788 clips, 5 s
annotations/train/<generator>.json   the gold labels (copy of the GitHub repository; captions and rubrics live there)

Generators: CogVideoX1.5-5B, HunyuanVideo-1.5 (480p), LTX-Video-2B and Wan2.2-T2V-A14B.

Usage

Code, the judge-training files (data/sft_train.jsonl, ...) and the training scripts are at https://github.com/chenyil6/phyRubric. Download this dataset into the repository so that the relative video paths resolve:

git clone https://github.com/chenyil6/phyRubric && cd phyRubric
hf download kittyhy/PhyRubric-Train --repo-type dataset --local-dir PhyRubric-Train
MODEL=Qwen/Qwen3-VL-8B-Instruct OUT=runs/sft bash scripts/train_sft.sh

The evaluation split is PhyRubric-Bench; the judge weights are PhyRubric-Judge.

The videos were generated by the listed models from our captions and are released for research use; each generator's own licence applies to its outputs.