Appendix B s1K details

本文为论文《s1：用简单方法实现测试时计算扩展》「论文原文」板块的对应章节，按 arXiv 官方全文逐字保留，层级与「中文全文翻译」板块一一对应。

B.1 s1K summary

Table 5: Summary of our dataset s1K. Token count measured by the Qwen-2.5 tokenizer. We prompt Claude to produce keywords given several questions from the domain.

B.2 Dataset composition for full 59K questions

Table 6: Composition of full 59K questions. Thinking and response lengths are measured in tokens using the Qwen2.5-32B-Instruct tokenizer (Qwen et al., 2024). In addition to excluding our evaluation benchmark, AIME24, we also exclude AIME questions from 2022-2023 as we use these 90 questions during our development stage of s1-32B.

B.3 s1K grading prompt

To grade whether an example is correct for our dataset selection in §2, we use the prompt in Figure 8.

Figure 8: Grading prompt.

B.4 s1K diversity selection

Algorithm 1 Two-stage sampling for s1K

Algorithm 1provides our algorithm for selecting data in our diversity selection stage. As mentioned in §2, we also include samples from some specific benchmarks we perceive as high-quality. None of the samples overlap with our final evaluation.

B.5 Decontamination

We filter all samples by checking for an 8-gram overlap between the selected examples and the evaluation benchmarks: MATH500, GPTQA Diamond, and AIME24. We exclude questions with more than an 8-gram overlap.