werty1248/SFT-Reasoning-Ko-Regenerated-HyperCLOVAX
收藏Hugging Face2025-12-12 更新2025-12-20 收录
下载链接：
https://hf-mirror.com/datasets/werty1248/SFT-Reasoning-Ko-Regenerated-HyperCLOVAX
下载链接
链接失效反馈官方服务：
资源简介：
该数据集包含两个配置：Natural_Reasoning_gpt_oss_120b_s1和YiSang_HighQuality，用于韩语推理任务。每个配置包含索引、指令、韩语推理（reasoning_ko）、答案、完成原因（finish_reason）和来源等特征。数据集用于训练模型执行韩语推理任务，部分案例涉及将英语问题翻译为韩语推理和答案。内容还讨论了使用此数据集进行训练的性能影响，特别是在韩语基准测试中表现提升，但在数学推理分数上除非使用原始英语推理否则会下降。
This dataset includes two configurations: Natural_Reasoning_gpt_oss_120b_s1 and YiSang_HighQuality, designed for Korean reasoning tasks. Each configuration features index, instruction, reasoning in Korean (reasoning_ko), answer, finish_reason, and source. The dataset is utilized to train models for performing reasoning tasks in Korean, with some instances involving English questions translated into Korean reasoning and answers. The discussion also covers the performance impacts of training with this dataset, highlighting improvements in Korean benchmarks but a decline in mathematical reasoning scores unless the original English reasoning is employed.
提供机构：
werty1248
