touristgpt/Codeforces-COTs

Codeforces Trace Dataset是一个基于9,413个Codeforces问题（难度范围800-3500）的合成推理轨迹数据集。每个问题包含结构化的错误路径探索、发现叙述和伪代码，旨在教模型如何思考竞争性编程问题，而不仅仅是解决它们。数据集捕捉完整的推理过程，包括：1. 错误路径——详细探索无效方法及其失败原因；2. 发现叙述——第一人称视角描述如何从失败方法转向正确方法，并逐步跟踪示例；3. 伪代码——为正确方法提供清晰、可实现的伪代码；4. 新问题——包含新竞赛的问题和Open-r1数据集的所有问题。数据集模拟专家思考方式，尝试想法、遇到障碍、从失败中提取见解并构建解决方案。数据集结构分为两个难度层级（≤1700和≥1800），包含22个字段，涵盖问题元数据和生成轨迹字段。生成管道使用DeepSeek V3.2的三阶段流程。数据集可用于SFT训练、过程监督、错误路径学习和课程学习。

Codeforces Trace Dataset is a synthetic reasoning-trace dataset built from 9,413 Codeforces problems (800–3500 rating). Each problem includes structured wrong-path explorations, discovery narratives, and pseudocode designed to teach models how to think through competitive programming problems, not just solve them. The dataset captures the full reasoning process, including: 1. Wrong paths—detailed explorations of approaches that dont work, including why they fail; 2. Discovery narrative—a first-person account of pivoting from failed approaches to the correct one, with worked examples traced step-by-step; 3. Pseudocode—clean, implementation-ready pseudocode for the correct approach; 4. Newer Questions—questions from newer contests and all questions from the Open-r1 dataset. The dataset models how an expert thinks—trying ideas, hitting walls, extracting insights from failures, and building toward the solution. The dataset structure has two tiers based on difficulty (≤1700 and ≥1800), with 22 columns covering problem metadata and generated trace fields. The generation pipeline uses a 3-stage process with DeepSeek V3.2. The dataset is intended for SFT training, process supervision, wrong-path learning, and curriculum learning.