SynLogic is a comprehensive synthetic logic reasoning dataset released in 2025 by the Hong Kong University of Science and Technology and MiniMax research team. The related paper results are:SynLogic: Synthesizing Verifiable Reasoning Data at Scale for Learning Logical Reasoning and Beyond", which aims to enhance the logical reasoning ability of large language models (LLMs) through reinforcement learning with verifiable rewards.

The dataset contains 35 diverse logical reasoning tasks and has automatic validation capabilities, making it well suited for reinforcement learning training.

Main Features

• 35 task types: including Sudoku, 24-point games, passwords, arrow mazes, arithmetic puzzles, etc.
• Verifiable Bonus: All samples have an automatic verifier for correctness checking
• Controllable difficulty: Each task has adjustable difficulty parameters
• Two versions: easy version (for 7B model) and difficult version (for 32B model)

Dataset Configuration

SynLogic-Easy

• Target: 7B parameter model
• Task: 27 missions
• sample: About 16,000 training instances

SynLogic-Hard

• Target：32B parameter model
• Task: All 35 tasks
• sample: About 33,000 training instances