Wisdom-math/wisdom-math

WISDOM数据集通过渐进式课程合成方法生成数学推理问题，并利用Deepseek Coder V2和GPT-4o进行回答生成。该数据集旨在通过从易到难的课程学习，逐步合成高质量的CoT（Chain-of-Thought）数据，以提升大语言模型在数学推理任务中的表现。训练过程中使用了Alpaca模板，并采用了10-gram哈希去重方法来避免数据污染。训练在88个NVIDIA A800 GPU上进行，使用了AdamW优化器和余弦学习率调度。

The WISDOM dataset generates mathematical reasoning questions through a progressive curriculum synthesis method and utilizes Deepseek Coder V2 and GPT-4o to generate responses. The dataset aims to gradually synthesize high-quality CoT (Chain-of-Thought) data from easy to hard through curriculum learning, enhancing the performance of large language models in mathematical reasoning tasks. The training process uses the Alpaca template and employs a 10-gram hash deduplication method to avoid data contamination. Training is conducted on 88 NVIDIA A800 GPUs, using the AdamW optimizer and a cosine learning rate schedule.

搜集汇总

背景与挑战

背景概述

The 'Wisdom-math/wisdom-math' dataset is a comprehensive collection of mathematical problems and solutions, designed to improve the mathematical reasoning of large language models through progressive curriculum learning. It includes a wide range of topics and is formatted in JSON, making it suitable for training and evaluating models on complex mathematical tasks.