Appendix B English web ablations

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。

Figure 4 shows the evaluation curves of ablation models trained on DCLM, FineWeb-Edu and their mix for 350B tokens.

Figure 4: Evaluation of models trained on FineWeb-Edu and DCLM for 350B tokens. FineWeb-Edu excels at knowledge and reasoning tasks, while DCLM demonstrates stronger performance on commonsense reasoning benchmarks. A 60/40 mixture of FineWeb-Edu and DCLM achieves balanced performance across all tasks.