metadata
license: mit
tags:
- benchmark
task_categories:
- multiple-choice
- question-answering
QUAD-Bench: A Lightweight AI Reasoning & Acuity Benchmark
QUAD-Bench is a compact, multiple-choice benchmark dataset designed for quick evaluation of Large Language Models (LLMs). The dataset contains 100 questions evenly distributed across 4 fundamental capabilities, requiring the model to select exactly one correct answer option (A, B, C, or D).
📊 Dataset Structure
The benchmark consists of 100 questions divided into 4 categories (25 questions each):
| Category | Description | Question Range |
|---|---|---|
| Logic | Logical deductions, conditional statements, syllogisms, and classic riddles. | id: 1–25 |
| Attention | Pattern recognition, string manipulation, character counting, and detail sensitivity. | id: 26–50 |
| Math | Basic arithmetic, algebra, probability, geometry, and calculus concepts. | id: 51–75 |
| Code | Syntax understanding, algorithm complexities, data structures, and tech fundamentals. | id: 76–100 |
📁 JSON Format
Each entry in dataset.json follows this simple schema:
{
"id": 1,
"category": "logic",
"question": "If all Bloops are Razzies and all Razzies are Lazzies, then all Bloops are definitely Lazzies.\nA) True\nB) False\nC) Cannot be determined\nD) None of the above",
"answer": "A"
}