还有一个重要的指标——准确率。伯克利函数调用排行榜 (BFCL) 是评估函数调用能力的标准基准。 Gemma 3 1B 的得分约为 31%，Llama 3.2 1B 约为 26%，两者未经微调的性能都很弱。由于 Gemma 3n 是通用型程序，因此未对其进行测试。Hammer 2.1 0.5B 没有公开数据，但其 1.5B 版本开箱即用的得分约为 73%——尽管它在 int8 内存中占用约 1.5GB 的空间，是 FunctionGemma（288MB）的 5 倍。
Negative net migration hasn’t definitively happened in the U.S. for almost a century, not since the Great Depression. The U.S. native-born population is also in free fall, well below the replacement rate needed to maintain population numbers in the long run. With fewer immigrant arrivals already dealing a blow to the workforce, the shrinkage of the American taxpayer base is set to further widen the country’s deficit and hurt its prospects for economic growth.
，更多细节参见91视频
What should admins do if certain people are causing problems?
in this area, too. It takes a combination of both tools to gain an advantage