Atarogic/BiasIG
收藏资源简介:
--- license: mit task_categories: - text-to-image language: - en tags: - bias - fairness - text-to-image - benchmark - social-bias - arxiv:2604.11934 pretty_name: BiasIG size_categories: - 10K<n<100K configs: - config_name: default data_files: - split: train path: prompts.jsonl --- # BiasIG BiasIG is a benchmark for auditing multi-dimensional social biases in text-to-image models. This dataset release contains the prompt suite from the [BiasIG GitHub repository](https://github.com/Astarojth/BiasIG), together with the released ground-truth statistics used by the evaluation code. - Paper: [BiasIG: Benchmarking Multi-dimensional Social Biases in Text-to-Image Models](https://arxiv.org/abs/2604.11934) - Code and benchmark repository: [Astarojth/BiasIG](https://github.com/Astarojth/BiasIG) - Evaluation backbone: [BIGBench/InternVL-4B-bench](https://huggingface.co/BIGBench/InternVL-4B-bench) ## What Is Included - 47,040 text-to-image prompts in `prompts.jsonl` - Ground-truth statistics and weighting files in `truth/` - Metadata fields derived from the original prompt layout: `setting`, `family`, `axis`, and `attribute_values` The prompt suite covers implicit and explicit evaluation settings across occupations, characteristics, and social relations. ## Data Fields Each prompt record contains: - `id`: stable row identifier in this export - `prompt`: text-to-image prompt - `source_file`: original path in the BiasIG repository - `line_number`: original line number - `setting`: `implicit` or `explicit` - `family`: `occupation`, `characteristic`, or `relation` - `axis`: protected or evaluation axis when available - `attribute_values`: explicit attribute values parsed from the filename when available ## Usage ```python from datasets import load_dataset dataset = load_dataset("Atarogic/BiasIG") ``` The released ground-truth files can be accessed directly from the repository under `truth/`. ## Citation If you use BiasIG, please cite the paper and repository. ```bibtex @article{luo2026biasig, title={BiasIG: Benchmarking Multi-dimensional Social Biases in Text-to-Image Models}, author={Luo, Hanjun and Huang, Zhimu and Huang, Haoyu and Deng, Ziye and Chen, Ruizhe and Li, Xinfeng and Liu, Zuozhu and Salam, Hanan}, journal={arXiv preprint arXiv:2604.11934}, year={2026}, url={https://arxiv.org/abs/2604.11934} } ```
BiasIG is a benchmark for auditing multi-dimensional social biases in text-to-image models.




