MiniMaxAI/SynLogic-7B
MiniMaxAI/SynLogic-7B is a 7.6 billion parameter logical reasoning model built on Qwen2.5-7B-Base with a 131072 token context length. It is fine-tuned using reinforcement learning on 27 diverse logical reasoning tasks, demonstrating strong generalization to mathematical problem-solving without explicit math training. This model excels in complex reasoning tasks, outperforming its base model on benchmarks like KOR-Bench and AIME 2024.
Loading preview...
SynLogic-7B: A Logical Reasoning Powerhouse
SynLogic-7B, developed by MiniMaxAI, is a 7.6 billion parameter model based on Qwen2.5-7B-Base, specifically engineered for advanced logical reasoning. It leverages reinforcement learning on a unique, verifiable dataset of 27 diverse logical tasks, including Sudoku and Game of 24, to achieve robust reasoning capabilities.
Key Capabilities & Features
- Enhanced Logical Reasoning: Significantly improves performance on logical reasoning benchmarks like KOR-Bench (achieving 48.1 vs. 38.6 for Qwen2.5-7B-Instruct).
- Mathematical Generalization: Demonstrates strong transfer learning to mathematical domains, scoring 10.0% on AIME 2024 without explicit math training.
- Verifiable Training Data: Utilizes a dataset where all training examples can be automatically verified, enabling effective and scalable reinforcement learning.
- Efficient Scale: Delivers strong performance in logical and mathematical reasoning with a compact 7B parameter count.
- Advanced Training: Trained using Group Relative Policy Optimization (GRPO) on 16k SynLogic-Easy samples, optimized for this model size.
Ideal Use Cases
- Complex Problem Solving: Suited for applications requiring intricate logical deduction and multi-step reasoning.
- Mathematical Assistance: Can be applied to mathematical problem-solving where logical inference is key, despite not being explicitly math-trained.
- Research & Development: Valuable for exploring advanced reasoning capabilities in smaller, efficient models.