Appendix G Long context evaluations

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 7 shows the evaluation results on the Needle in the Haystack benchmark.

Figure 7: Needle in the Haystack evaluation of SmolLM2 with 8192 context length.

Table 11 shows the evaluation results on the HELMET benchmark.

Table 11: Evaluation results of the base models on the HELMET benchmark using 8k maximum input length.