dataset_info:
features:
- name: id
dtype: string
- name: task_type
dtype: string
- name: image
dtype: binary
- name: question
dtype: string
- name: answer
dtype: string
- name: choices
dtype: string
- name: metadata
dtype: string
splits:
- name: test
num_bytes: 740966
num_examples: 500
download_size: 740966
dataset_size: 740966
configs:
- config_name: default
data_files:
- split: test
path: data/vlbenchy.parquet
Dataset Summary
VLbenchy is a 500-sample procedural vision-language benchmark for evaluating the fundamental visual understanding capabilities of VL models across 10 distinct task types.
Dataset Structure
Each sample contains a procedurally generated 336×336 PNG image paired with a natural-language question, a ground-truth answer, and multiple-choice options. All images are unique — generated with randomised shapes, colors, sizes, and layouts.
Data Fields
id (string): Unique sample identifier (e.g. "vlb_0042")
task_type (string): One of 10 task categories — color_recognition, shape_counting, ocr, spatial_reasoning, size_comparison, object_presence, color_counting, grid_pattern, odd_one_out, background_color
image (binary): Raw PNG image bytes (336×336 px)
question (string): Natural-language question about the image
answer (string): Ground-truth answer string (exact-match scoreable)
choices (string): JSON-encoded list of 2–4 multiple-choice options
metadata (string): JSON-encoded dict with generation parameters (shapes, colors, counts, etc.)
Task Types
| Task | Samples | Description |
|---|---|---|
| color_recognition | 50 | Identify the color of a shape |
| shape_counting | 50 | Count how many of a given shape are visible |
| ocr | 50 | Read a number rendered in the image |
| spatial_reasoning | 50 | Determine positional relationship between two shapes |
| size_comparison | 50 | Identify which of two shapes is larger |
| object_presence | 50 | Detect whether a specific shape is present |
| color_counting | 50 | Count the number of distinct colors among shapes |
| grid_pattern | 50 | Count rows or columns in a grid of shapes |
| odd_one_out | 50 | Identify the shape that differs from the others |
| background_color | 50 | Identify the background color of the image |