stepfun-ai/PaCoRe-Train-8k

PaCoRe（并行协调推理）数据集是一个用于训练模型执行推理任务的高质量语料库，包含开源数学、公共数学竞赛、合成数学和代码等数据。该数据集支持并行探索轨迹和消息传递架构，旨在突破模型上下文限制，大规模扩展测试时计算。数据集以`list[dict]`格式提供，每个条目代表一个训练实例，包含原始问题/提示消息、生成的响应列表（轨迹）和用于正确性评估的验证答案。

The PaCoRe (Parallel Coordinated Reasoning) dataset is a high-quality training corpus for models to perform reasoning tasks, including opensource math, public math contests, synthetic math, and code. The dataset supports parallel exploration trajectories and a message-passing architecture, aiming to break the model context limitation and massively scale test-time compute. The data is provided as a `list[dict]`, where each entry represents a training instance containing the original problem/prompt messages, a list of generated responses (trajectories), and a verifiable answer for correctness evaluation.