nphearum/Code-Reasoning-4k

Code-Reasoning-4k是一个专注于代码推理任务的数据集，包含38,140个样本，远超最初命名的4k规模。数据集设计用于支持代码生成和调试、算法推理、数学问题解决和一般指令任务的模型训练和评估。尽管目标是代码、数学和一般推理的比例为70%、20%和10%，但实际上代码占比高达91.8%，数学和一般推理分别占5.4%和2.9%。数据集的结构包括提示、响应和可选的推理轨迹，适用于监督微调、推理链学习和代码补全任务。数据集的主要来源包括codex-2m、crownelius、teichai和nohurry，其中codex-2m占主导地位。数据集在代码推理任务上表现强劲，但在平衡多领域推理基准测试上可靠性较低。

Code-Reasoning-4k is a curated dataset for code-centric reasoning tasks, combining programming problems, mathematical reasoning, and general instruction-following samples. Despite the name, the dataset currently contains 38,140 samples, reflecting a significant expansion beyond the original 4k scale. The dataset is designed to support training and evaluation of models on code generation and debugging, algorithmic reasoning, mathematical problem solving, and light general reasoning / instruction tasks. The intended ratio of code, math, and general reasoning was 70%, 20%, and 10%, but the actual distribution is 91.8% code, 5.4% math, and 2.9% general. Each entry typically contains a prompt, response, and optional reasoning traces, supporting supervised fine-tuning, reasoning chain learning, and code completion tasks. The dataset is heavily skewed toward code, making it strong for code reasoning training but less reliable for balanced multi-domain reasoning benchmarks.