Bespoke Stratos 17k 推理任务数据集

Bespoke-Stratos-17k 是一个专为推理任务设计的高质量数据集，由 Bespoke Labs 团队于 2025 年开发，相关 Blog 为「Bespoke-Stratos: The unreasonable effectiveness of reasoning distillation」。该数据集通过改进伯克利的 Sky-T1 数据管道，并利用 DeepSeek-R1 的蒸馏数据生成，旨在为训练高性能推理模型提供支持。数据集包含问题、推理轨迹和答案，覆盖代码、数学和科学谜题等多个领域。通过使用 Bespoke Curator 工具，仅用 1.5 小时即可生成高质量的推理数据集，成本控制在 800 美元左右。该数据集采用 DeepSeek-R1 作为教师推理模型，无需额外格式化步骤，简化了数据生成流程。此外，通过 gpt-4o-mini 过滤错误的数学解决方案，显著提高了正确数学解决方案的保留率，从 25% 提升至 73% 。

Bespoke-Stratos-17k is a high-quality dataset specifically designed for reasoning tasks, developed by the Bespoke Labs team in 2025, with the accompanying blog post titled "Bespoke-Stratos: The unreasonable effectiveness of reasoning distillation". This dataset is constructed by improving Berkeley's Sky-T1 data pipeline and leveraging distillation data generated by DeepSeek-R1, aiming to support the training of high-performance reasoning models. The dataset includes questions, reasoning trajectories and answers, covering multiple domains such as code, mathematics and scientific puzzles. By utilizing the Bespoke Curator tool, this high-quality reasoning dataset can be generated in merely 1.5 hours, with the cost controlled at approximately $800. The dataset adopts DeepSeek-R1 as the teacher reasoning model, eliminating the need for additional formatting steps and simplifying the data generation workflow. Furthermore, erroneous mathematical solutions are filtered via gpt-4o-mini, which significantly boosts the retention rate of correct mathematical solutions from 25% to 73%.

搜集汇总

背景与挑战

背景概述

Bespoke Stratos 17k是由Bespoke Labs于2025年开发的推理任务数据集，通过改进Sky-T1管道并利用DeepSeek-R1蒸馏生成，覆盖代码、数学和科学谜题领域。它包含17,000条数据，用于训练高性能推理模型，在基准测试中表现优异，且通过过滤机制提高了数学解决方案的正确率。