VlmBias — Vision-Language Models Are Biased
VLMbias is an evaluation game designed to highlight visual reasoning errors caused by biases stored in vision-language models.
Image-question pairs in JSON, classified by 7 domains (logos, animals, games, illusions, etc.)
MIT
Description
VlmBias is a visual benchmark that aims to test the real robustness of vision-language models (VLMs) in the face of subtle modifications in known images. It covers 7 categories: animals, logos, flags, flags, flags, chess, board games, optical illusions, and patterned grids. Each image is associated with a question, often counting or identifying anomalies, and some contain textual decoys to exacerbate biases.
What is this dataset for?
- Measuring the cognitive limits of VLMs in situations of subtle visual recognition
- Develop models that are more robust to semantic bias or over-learning
- Test counting systems, pattern recognition, or change detection
Can it be enriched or improved?
Absolutely. It is possible to add new visual categories, to increase the number of counterfactuals per domain, or to add other types of questions (VQA, justification, etc.). Multilingual tests or cultural variants could also increase its reach.
🔎 In summary
🧠 Recommended for
- VLM evaluators
- NLP/CV robustness researchers
- Fairness/bias teams
🔧 Compatible tools
- BLIP
- Lava
- OpenFlamingo
- CLIP
- Hugging Face Transformers
💡 Tip
Test your models with and without explicit instructions to see if this reduces bias on counterfactual images.
Frequently Asked Questions
Can this dataset be used for training?
It is primarily designed for evaluation, but some parts can be used to train robust models or anomaly detectors.
What types of biases are studied with this corpus?
Mainly the biases associated with excessive semantic memory, where the model favors what it “knows” rather than what it “sees”.
Are the images realistic or artificial?
They are realistic but subtly modified, making them perfect for testing the real visual perception of models.
