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。
推进中国式现代化，短板在农业农村，潜力也在农业农村。
，这一点在safew官方下载中也有详细论述
One reason for these price drops is that the value of its most expensive chase cards has fallen since this Scarlet and Violet set expansion launched in March 2025. Be that as it may, there are still some nice cards worth pulling (that are also available as singles on TCGplayer) — such as Lillie’s Clefairy ex, Salamence ex, and N’s Zoroark ex.
Factorized embed, rotation Q (2 angles), tied embed+V dir, rank-1 MLP, parabolic head, sinusoidal PE (period 11)