MiniMax-AI/SynLogic

MiniMax-AI/SynLogic

Description: [NeurIPS 2025] The official repo of SynLogic: Synthesizing Verifiable Reasoning Data at Scale for Learning Logical Reasoning and Beyond

Language: Python

License: MIT

Stars: 203

Forks: 22

Open issues: 1

Created: 2025-05-23T14:54:08Z

Pushed: 2025-07-07T06:03:03Z

Default branch: main

Fork: no

Archived: no

README:

SynLogic

This repository contains the code and data for SynLogic, a comprehensive logical reasoning data synthesis framework that generates diverse, verifiable reasoning data at scale. Our work addresses the critical gap in high-quality logical reasoning training data for developing general reasoning capabilities in Large Language Models (LLMs).

News

• [2025/07/07] :fire: We provide guidance for using SynLogic data to conduct RL training with the Verl framework. See [docs/training_guidance.md](docs/training_guidance.md) for details.

• [2025/06/03] :fire: Trained Models Released on Hugging Face:

→ MiniMaxAI/SynLogic Collection

• [2025/05] :fire: We are excited to release the resources for the paper "SynLogic: Synthesizing Verifiable Reasoning Data at Scale for Learning Logical Reasoning and Beyond"

📋 Overview

SynLogic is a data synthesis framework and comprehensive dataset that:

• 📊 Covers 35 diverse logical reasoning tasks including Sudoku, Game of 24, Cipher, Arrow Maze, and more
• 🎯 Enables controllable difficulty synthesis with adjustable parameters for each task
• ✅ Provides verifiable rewards through rule-based verification for RL training
• 🚀 Achieves SOTA performance among open-source datasets, surpassing DeepSeek-R1-Distill-Qwen-32B by 6 points on BBEH

Key Features

• Scalable Data Generation: Unlimited synthesis with controllable difficulty levels
• Comprehensive Task Coverage: 35 distinct logical reasoning tasks with custom generators and verifiers
• RL-Ready: All examples can be verified by simple rules, making them ideal for reinforcement learning
• Cross-Domain Transfer: Strong generalization to mathematical and coding domains

🚀 Quick Start

Installation

git clone https://github.com/MiniMax-AI/SynLogic.git
cd SynLogic
pip install -r requirements.txt

Generate Sample Data (Arrow Maze Example)

# Quick example with Arrow Maze
bash games/tasks/arrow_maze/run.sh

# Or with custom parameters
python scripts/arrow_maze.py \
--num_of_data 1000 \
--width 5 \
--height 5 \
--arrow_fill_rate_min 0.3 \
--arrow_fill_rate_max 0.9

For RL Training

We provide comprehensive guidance for using SynLogic data in reinforcement learning training. See [docs/training_guidance.md](docs/training_guidance.md) for detailed instructions on integrating SynLogic with the Verl framework.

:rocket: Resources

Datasets

| Dataset Name | Description | Link | |:------------:|:------------|:----:| | SynLogic | Dataset | 🤗 HuggingFace |

Models

| Model Name | Description | Link | |:-----------|:------------|:----:| | SynLogic-7B | 7B model trained on SynLogic | 🤗 HuggingFace | | SynLogic-32B | 32B model trained on SynLogic | 🤗 HuggingFace | | SynLogic-Mix-3-32B | 32B model trained on mixed SynLogic, Math, Coding data | 🤗 HuggingFace |

🔄 Future Updates

We are actively working on expanding SynLogic with more logical reasoning tasks.

Follow our repository for the latest updates and releases!

Citation

Please cite our paper if you find our work helpful:

@misc{liu2025synlogic,
title={SynLogic: Synthesizing Verifiable Reasoning Data at Scale for Learning Logical Reasoning and Beyond},
author={Junteng Liu and Yuanxiang Fan and Zhuo Jiang and Han Ding and Yongyi Hu and Chi Zhang and Yiqi Shi and Shitong Weng and Aili Chen and Shiqi Chen and Yunan Huang and Mozhi Zhang and Pengyu Zhao and Junjie Yan and Junxian He},
year={2025},
eprint={2505.19641},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2505.19641},
}