7 Conclusion

本文为论文《SmolLM2：紧凑高效的小语言模型家族》「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。

SmolLM2 advances the state-of-the-art for open small LMs through a combination of careful dataset curation and multi-stage training. Our approach highlights the critical role of high-quality, specialized datasets in enabling smaller models to achieve strong performance across a variety of benchmarks. The development of FineMath, Stack-Edu, and SmolTalk addressed limitations in existing public datasets, improving capabilities in reasoning, mathematics, and instruction-following tasks. To support future research and development, we release SmolLM2 alongside the datasets and code used in its training. These resources provide a comprehensive foundation for training performant small language models, making them accessible to a broader range of researchers and applications.