reasoning-data-numina
收藏资源简介:
该数据集是一个数学问题求解记录数据集,包含两个独立配置(minimax和qwen)。每个配置包含50,000个训练样本,记录了大型语言模型在解决数学问题时的多轮尝试过程。数据集的核心特征包括:问题索引、最终正确性判断、问题文本(仅minimax配置包含)、预期答案、正确的推理过程、正确答案、推理和解决方案的token长度统计、总尝试次数。每个样本还包含详细的尝试记录列表,记录每次尝试的具体答案、尝试编号、正确性判断、推理过程、解决方案以及对应的token长度。该数据集特别适用于数学推理能力评估、大语言模型错误模式分析、多步推理过程研究以及数学问题求解算法的开发与优化。minimax配置约2.44GB,qwen配置约3.61GB。
This dataset is a collection of mathematical problem-solving records, containing two independent configurations (minimax and qwen). Each configuration includes 50,000 training samples, documenting the multi-round attempts of large language models in solving mathematical problems. Core features of the dataset include: problem index, final correctness judgment, problem text (only included in the minimax configuration), expected answer, correct reasoning process, correct answer, token length statistics for reasoning and solutions, and total number of attempts. Each sample also contains a detailed list of attempt records, documenting the specific answer, attempt number, correctness judgment, reasoning process, solution, and corresponding token length for each attempt. The dataset is particularly suitable for mathematical reasoning ability evaluation, error pattern analysis of large language models, multi-step reasoning process research, and the development and optimization of mathematical problem-solving algorithms. The minimax configuration is approximately 2.44GB, and the qwen configuration is approximately 3.61GB.
数据集概述
数据集名称
reasoning-data-numina
数据集地址
https://huggingface.co/datasets/annakosovskaia/reasoning-data-numina
数据集配置
该数据集包含两个配置(config):
-
minimax
- 训练集(train):50,000 个样本
- 数据大小:约 2.44 GB(2,438,626,040 字节)
- 下载大小:约 2.15 GB(2,145,427,043 字节)
- 数据文件路径:
minimax/train-*
-
qwen
- 训练集(train):50,000 个样本
- 数据大小:约 3.63 GB(3,625,462,832 字节)
- 下载大小:约 3.08 GB(3,077,515,388 字节)
- 数据文件路径:
qwen/train-*
数据特征
每个配置包含相同的特征字段:
| 字段名 | 数据类型 | 说明 |
|---|---|---|
| problem_index | int64 | 问题索引编号 |
| is_correct | bool | 是否正确 |
| question | string | 问题内容 |
| expected_answer | string | 期望答案 |
| correct_reasoning | string | 正确的推理过程 |
| correct_answer | string | 正确答案 |
| tok_reasoning | int64 | 推理过程的 token 数量 |
| tok_solution | int64 | 解决方案的 token 数量 |
| total_attempts | int64 | 总尝试次数 |
| attempts | list | 每次尝试的详细信息(列表) |
attempts 列表中的字段:
| 字段名 | 数据类型 | 说明 |
|---|---|---|
| answer | string | 该次尝试的答案 |
| attempt_number | int64 | 尝试编号 |
| is_correct | bool | 该次尝试是否正确 |
| reasoning | string | 该次尝试的推理过程 |
| solution | string | 该次尝试的解决方案 |
| tok_reasoning | int64 | 该次尝试推理的 token 数量 |
| tok_solution | int64 | 该次尝试解决方案的 token 数量 |




