VLM Reality Check is a large-scale, automatically generated dataset designed to evaluate the causal reasoning and bias resistance of Vision-Language Models (VLMs). It features nearly 100,000 challenges across 14 bias types, utilizing counterfactual image transforms to create minimal-edit contrastive pairs.

Dataset Summary

Total Challenges: 95,317
Source Images: 1,868 (from Kaggle/ECCV v3 pipeline)
Counterfactual Images: 27,728
Bias Types: 14 (Spatial, Compositional, Counting, Cultural, etc.)
Languages: 5 (English, Spanish, Chinese, Hindi, Arabic)
Total Multilingual Instances: 476,585
Size: ~3.9 GB

Bias Types and Distribution

Bias Type	Count
Temporal Reasoning	12,001
Counting	9,932
Cultural Visual Bias	9,281
Texture	8,817
Spatial Relations	8,691
Physical Plausibility	8,433
Temporal Consistency	8,172
Text in Image	7,545
Occlusion Gradient	6,176
Scale Invariance	6,064
Typography Conflict	6,063
Compositional Binding	2,191
Compound	1,025
Spurious Correlation	926

Difficulty Distribution

Easy: 34,195
Medium: 38,902
Hard: 22,220

Dataset Structure

images/: Original source images.
counterfactuals/: Generated counterfactual images (split into subdirectories 0, 1, 2 to comply with directory limits).
annotations/: YOLO-based detections and metadata for source images.
challenges/: The core challenge JSONL files.
translations/: Multilingual versions of the challenges.
configs/: Generation configurations.
final_dataset.jsonl: The complete dataset in a single file.

How to Use

Each entry in final_dataset.jsonl contains:

challenge_id: Unique identifier.
bias_type: The category of visual reasoning tested.
image_a_id / image_b_id: References to images in images/ or counterfactuals/.
question_translated: The VQA question.
correct_answer: The ground truth.
distractor_answers: Incorrect but plausible options.

License

This dataset is released under the Apache 2.0 License.

Downloads last month: 7