Key-Point-Driven Data Synthesis with its Enhancement on Mathematical Reasoning
This addresses the problem of limited reasoning-focused datasets for researchers and practitioners in AI, enabling improved mathematical reasoning in LLMs, though it is incremental as it builds on existing data synthesis methods.
The paper tackles the scarcity of high-quality training data for mathematical reasoning in large language models by proposing Key-Point-Driven Data Synthesis (KPDDS), which generates over 800K question-answer pairs, leading to a fine-tuned model achieving 87.0% accuracy on GSM8K and 58.3% on MATH, outperforming competitors including GPT-4.
Large language models (LLMs) have shown great potential in complex reasoning tasks, yet their performance is often hampered by the scarcity of high-quality and reasoning-focused training datasets. Addressing this challenge, we propose Key-Point-Driven Data Synthesis (KPDDS), a novel data synthesis framework that synthesizes question-answer pairs by leveraging key points and exemplar practices from authentic data sources. KPDDS ensures the generation of novel questions with rigorous quality control and substantial scalability. As a result, we present KPMath, an extensive synthetic dataset tailored for mathematical reasoning, comprising over 800K question-answer pairs. Utilizing KPMath and augmenting it with additional reasoning-intensive corpora, we create the comprehensive KPMath-Plus dataset. The Qwen1.5-72B model, fine-tuned on KPMath-Plus, achieves 87.0% PASS@1 accuracy on GSM8K and 58.3% on MATH, surpassing competitors in the 7B to 70B range and best commercial models like GPT-4 across multiple math reasoning datasets.