OmniFair: A Unified Fairness Benchmark Across Tasks and Modalities

OmniFair is the first fairness benchmark with instance-level cross-task and cross-modal alignment. It introduces the Bias Semantic Unit (BSU), a task- and modality-independent representation of bias, and uses it to construct:

• BiasAtlas: a semantic space of 416K unique BSUs distilled from 143 fairness benchmarks and 11M multilingual news articles.
• OmniFair Benchmark: 46,125 evaluation instances spanning 5 task paradigms x 3 modalities (text, image, audio).

Key Findings

Evaluating 15 state-of-the-art MLLMs reveals:

• Fairness is not a unified capability: cross-task consistency is near random (Cohen's kappa = 0.02), while cross-modal consistency is moderate (kappa = 0.28).
• Semantic rankings are stable across models (cross-model Spearman rho = 0.82 for protected attributes), pointing to shared alignment deficiencies.
• When models discriminate, bias is overwhelmingly stereotypical in direction.

Dataset Structure

OmniFair-release/
├── BiasAtlas/
│   └── bsu.jsonl                    # 416,690 unique Bias Semantic Units
├── benchmark/
│   ├── text/                        # 5 paradigms x 3,075 BSUs = 15,375 samples
│   │   ├── open_generation.jsonl
│   │   ├── multiple_choice_qa.jsonl
│   │   ├── detection.jsonl
│   │   ├── continuation.jsonl
│   │   └── pairwise_comparison.jsonl
│   ├── image/                       # 15,375 samples + 6,150 images
│   │   ├── *.jsonl
│   │   └── images/                  # target + contrast images per BSU
│   └── audio/                       # 15,375 samples + media
│       ├── *.jsonl
│       ├── content_audio/           # 6,150 spoken-content clips (mp3)
│       └── acoustic_audio/          # 768 timbre clips for age/gender (wav)
├── inference/                       # Raw model outputs
│   ├── text/{text_thinking,text_nothink}/
│   ├── image/{image_thinking,image_nothink}/
│   └── audio/
└── results/                         # LLM-as-Judge evaluations
    ├── text/{text_thinking,text_nothink}/
    ├── image/{image_thinking,image_nothink}/
    └── audio/

BSU Schema

Each Bias Semantic Unit is a 5-dimensional representation:

Field	Description	Example
context	Deployment scenario	healthcare
protected_attribute	Comparison axis	race_ethnicity
target_group	Social group	Black and Latino communities
bias_expression	Stereotypical association	vulnerable to severe health outcomes
harm_category	Potential consequence	physical_violence

Task Paradigms

Abbr	Paradigm	Description
OG	Open Generation	Free-form response to a biased context
MCQ	Multiple-Choice QA	Select an answer; "Cannot be determined" is correct
DT	Detection	Binary bias detection
CT	Continuation	Complete a sentence or text prefix
PC	Pairwise Comparison	Compare counterfactual inputs differing in group identity

Evaluated Models

Closed-source (5): GPT-5.5, Claude-Sonnet-4.6, Gemini-3.1-Pro, Doubao-Seed-2.0, Qwen3.5-Omni-Plus

Open-source (10): DeepSeek-V4, Kimi-K2.6, GLM-4.6V, Qwen3.5 (4B/9B/27B/35B-A3B/122B-A10B/397B-A17B), Qwen3-Omni-30B

Usage

import json

# Load BiasAtlas
with open("BiasAtlas/bsu.jsonl") as f:
    bias_atlas = [json.loads(line) for line in f]

# Load a benchmark split
with open("benchmark/text/open_generation.jsonl") as f:
    og_samples = [json.loads(line) for line in f]

Citation

If you find this work useful, please cite our paper:

@article{ding2026omnifair,
  title={OmniFair: A Unified Fairness Benchmark Across Tasks and Modalities},
  author={Ding, Yifan and Feng, Yunhao and Li, Zhiyan and Zheng, Baihui and Gao, Yifeng and Li, Zihan and Wang, Guoyu and Ma, Xingjun and Jiang, Yu-Gang},
  journal={arXiv preprint},
  year={2026}
}

License

This dataset is released under CC BY 4.0.

The stereotypical content included in this dataset is intended solely for fairness evaluation and should not be interpreted as endorsement of any social group.