FIRM-Video-SFT-90K
This repository releases the 90K SFT data for FIRM-Video.
The dataset covers three key evaluation dimensions:
- Instruction Following (IF): whether the generated video accurately follows the text prompt.
- Visual Quality (VQ): perceptual and technical quality, including clarity, sharpness, artifacts, flicker, and overall visual fidelity.
- World Coherence (WC): whether the video is coherent with commonsense, temporal consistency, physical plausibility, and object/character structure.
For further details, please refer to the following resources:
- Paper:
- GitHub: https://github.com/VisionXLab/FIRM-Reward
- Models&Benchmark: https://huggingface.co/collections/VisionXLab/firm-reward
Citation