TACO-Cobalt
收藏资源简介:
TACO-Cobalt 是一个经过清理的代码生成数据集,源自编程竞赛网站爬取的 TACO 数据集。由于原始数据存在噪声,该数据集基于 TACO-verified 子集进一步清理,最终包含 6,103 个有效任务,每个任务至少包含 8 个测试用例。每个任务的测试用例根据 Qwen2.5-Coder-7B-Instruct 模型在 16 次尝试中通过的可能性进行难度排序,并选择四个最简单的测试用例作为公开测试集(平局时随机选择),其余作为隐藏测试集。此外,从五个标注的难度级别中随机选取 50 个示例,形成包含 250 个示例的验证集,其余 5,853 个示例构成训练集。该数据集适用于代码生成及相关研究任务。
TACO-Cobalt is a curated code generation dataset derived from the TACO dataset scraped from programming competition websites. Due to the noise inherent in the original raw data, this dataset is further cleaned using the TACO-verified subset, ultimately resulting in 6,103 valid tasks, each containing at least 8 test cases. For each task, test cases are ranked by difficulty based on the pass probability of the Qwen2.5-Coder-7B-Instruct model after 16 inference attempts. Four of the simplest test cases are selected as the public test set (randomly chosen in the event of tied difficulty scores), with the remaining test cases reserved as the hidden test set. Additionally, 50 examples are randomly sampled from each of the five annotated difficulty levels, creating a validation set with a total of 250 examples. The remaining 5,853 examples constitute the training set. This dataset is suitable for code generation and related research tasks.
TACO-Cobalt 数据集概述
数据集基本信息
- 数据集名称:TACO-Cobalt
- 许可证:MIT
- 来源:基于编程竞赛网站爬取数据清理而成
数据集描述
TACO-Cobalt 是从编程竞赛网站爬取的 TACO 代码生成数据集的清理版本。由于原始爬取数据存在噪声,本数据集从 TACO-verified 子集开始,经过进一步清理,最终包含 6,103 个有效任务,每个任务至少包含 8 个测试用例。
数据处理与划分
- 测试用例难度排序:针对每个任务,根据 Qwen2.5-Coder-7B-Instruct 模型在 16 次尝试中通过测试用例的可能性对测试用例难度进行排序。
- 公开与隐藏划分:
- 选择四个最简单的测试用例作为公开拆分,用于测试时交互(平局时随机选择)。
- 其余测试用例作为隐藏拆分。
- 数据集划分:
- 从五个标注难度级别中随机选择 50 个示例,组成包含 250 个示例的验证集。
- 其余 5,853 个示例包含在训练集中。
相关资源
- 原始论文:https://arxiv.org/abs/2312.14852
- GitHub 仓库:https://github.com/OSU-NLP-Group/cobalt
- 基础数据集:https://huggingface.co/datasets/likaixin/TACO-verified
引用信息
如果使用本数据集,请引用原始论文: bibtex @misc{li2023tacotopicsalgorithmiccode, title={TACO: Topics in Algorithmic COde generation dataset}, author={Rongao Li and Jie Fu and Bo-Wen Zhang and Tao Huang and Zhihong Sun and Chen Lyu and Guang Liu and Zhi Jin and Ge Li}, year={2023}, eprint={2312.14852}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2312.14852}, }




