Sign in to view source links and access this dataset
Description
56,475 synthetic visual scenes for training, plus 3,344 for testing, designed for Zendo-style rule learning. Each scene is labeled as positive or negative based on a logical rule expressed in natural language and Prolog, created by author sophia1ch. The dataset was last updated on June 4, 2026.
Use Cases
Train visual reasoning models based on synthetic scenes that follow or violate logical rules.
Benchmark neuro-symbolic AI systems based on the combination of image data and formal Prolog rule representations.
Develop models for multimodal rule induction based on natural language descriptions paired with visual examples.
Strengths
Contains 56,475 training scenes and 3,344 test scenes, providing a substantial scale for model development.
Includes 3,439 distinct logical rules, offering variety in reasoning tasks.
Provides multimodal data with images, natural language rules, and Prolog queries for each scene.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count for the full dataset is unknown, which may limit suitability assessment.
Data is synthetic, which may limit real-world applicability.
Provenance
Source
huggingface
Collection Method
Synthetically generated.
Freshness
Last updated 2026-06-04 06:50:02; freshness should be verified.
License is unknown; terms of use must be verified before application.