name: VLbenchy
description: VLbenchy is a 500-sample procedural vision-language benchmark for evaluating VL models on fundamental visual understanding skills. It covers 10 distinct task types including color recognition, shape counting, OCR, spatial reasoning, size comparison, object presence detection, color counting, grid pattern analysis, odd-one-out identification, and background color recognition. All samples are generated with randomised layouts ensuring no two images are identical.
evaluation_framework: vlbenchy

tasks:
  - id: VLbenchy
    config: default
    split: test