Dataset Card for CSL-Clinic

CSL-Clinic is a gloss-annotated Chinese Sign Language dataset for sign language understanding in the clinical and healthcare domain. It was introduced in Variational Sign Language Translation, published in the International Journal of Computer Vision.

Public Release and Full-Dataset Access

This Hugging Face repository publicly distributes only the 500-example test split. The complete dataset contains 5,972 examples; its train and dev splits are available to researchers upon request.

Split	Examples	Availability
train	5,000	Available by email request
dev	472	Available by email request
test	500	Public in this repository
Total	5,972

To request the complete dataset, follow the instructions in the GitHub repository and email the corresponding author at ydchen@xmu.edu.cn.

Dataset Structure

test/
├── metadata.csv
└── video/
    ├── example_0001.mp4
    └── ...

The public split contains 500 MP4 videos and one annotation row per video.

Data Fields

• file_name: relative path to the video inside the test directory;
• gloss: segmented Chinese Sign Language gloss sequence;
• text: corresponding Chinese natural-language sentence.

Intended Uses

CSL-Clinic supports research on continuous sign language recognition, sign language translation, gloss-to-text translation, multimodal learning, and accessible medical communication.

Out-of-Scope Uses

The dataset must not be used as a substitute for professional medical advice, diagnosis, emergency communication, or qualified human interpretation. Models trained or evaluated on this dataset may produce consequential recognition or translation errors.

Limitations and Privacy

• The dataset may not cover every Chinese Sign Language variant, signer style, medical specialty, or recording condition.
• Performance on CSL-Clinic may not generalize to real clinical environments.
• Video quality, signer identity, background, and recording conditions may affect model behavior.
• Users should not attempt to identify signers or infer private personal attributes.
• Non-public portions of the dataset must not be redistributed.

Citation

@article{zhao2026variational,
  title   = {Variational Sign Language Translation},
  author  = {Zhao, Rui and Zhang, Liang and Fu, Biao and Zhang, Ruiquan and Chen, Yidong and Shi, Xiaodong},
  journal = {International Journal of Computer Vision},
  volume  = {134},
  pages   = {408},
  year    = {2026},
  doi     = {10.1007/s11263-026-02978-x},
  url     = {https://doi.org/10.1007/s11263-026-02978-x}
}

Links

• Project repository
• Paper
• Implementation