CLEVR-MATE

MATE-like cross-modal entity linking dataset generated from CLEVR-style scenes for probe training.

Variants

Config	Renderer	Scenes	Samples	Strict	Description
2d	Pillow	3,000	18,000	No	Fast 2D shape rendering
pyrender	pyrender	3,000	18,000	No	Offscreen 3D rendering
blender	Blender Cycles	3,000	18,000	No	Photorealistic ray-traced rendering
2d_strict	Pillow	3,000	18,000	Yes	2D, strict cross-modal
pyrender_strict	pyrender	3,000	18,000	Yes	3D, strict cross-modal
blender_strict	Blender Cycles	3,000	18,000	Yes	Photorealistic, strict cross-modal

MATE-consistent vs Strict

• MATE-consistent (default): Matches MATE's convention — only the pointer/target visual attribute (color or shape) is stripped from the scene JSON. The other visual attribute remains, which may allow a "bridging shortcut" (e.g., if color is the pointer, shape is still in the JSON and visible in the image, so the model could match objects via shape alone without true cross-modal binding).

• Strict: Both color AND shape are stripped from the scene JSON, forcing the model to rely on non-visual attributes (name, size, rotation, 3d_coords) for entity linking. This eliminates the bridging shortcut and tests true cross-modal binding ability.

The strict variants share the same images as their non-strict counterparts — only the samples.jsonl (scene JSON filtering) differs.

Usage

from datasets import load_dataset

# Load a specific variant
ds = load_dataset("qinglinhou/CLEVR-MATE", "blender")

# Load the strict variant
ds_strict = load_dataset("qinglinhou/CLEVR-MATE", "blender_strict")

Fields

• id: unique hex identifier
• image_path: relative path to the scene image (e.g. images/scene_000042.png)
• scene: filtered scene JSON (Python repr format, visual attributes stripped)
• question: cross-modal question (e.g. "What is the name of the yellow colored object?")
• answer: gold answer
• task: img2data or data2img
• object_count: number of objects in the scene (3-8)
• pointer_attribute: attribute used to identify the object
• target_attribute: attribute to find

Task Types

16 unique (task, pointer, target) combinations matching MATE cross_modal format:

• img2data: pointer in {color, shape} -> target in {name, rotation, 3d_coords, size}
• data2img: pointer in {name, 3d_coords, rotation, size} -> target in {color, shape}