Appendix F Post-training

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。

Table 9 shows the final composition of SmolTalk dataset.

Table 9: Composition of the SmolTalk dataset. The total dataset contains 1.1M instruction-response pairs from different data sources.

Table 10 shows the performance after training on the different components of SmolTalk we consider. The top section compares the results of fine-tuning SmolLM2 base on different instruction datasets, while the bottom section evaluates the impact of adding 20% specialized math data to a base mixture of 80% MagPie-Ultra$\overset{+}{}$ during the SFT. The last row, SmolLM2-SFT, represents the final SFT checkpoint of SmolLM2 before DPO, trained for two epochs on the full SmolTalk dataset.

Table 10: Performance on instruction-tuning datasets. MagPie-Ultra$\overset{+}{}$ refers to MagPie-Ultra combined with Smol-Constraints, Smol-Rewrite, and Smol-Summarization. MagPie-Pro-MT is multi-turn while MagPie-Pro is the single turn version. All comparisons were performed by fine-tuning the SmolLM2 base model on each dataset for 1 epoch. SmolLM2-SFT†, the final supervised fine-tuned version of SmolLM2, was trained for 2 epochs on SmolTalk.