知识卡片
准确率与资源成本联合报告
某评估中的推理模型可能答对更多题,同时花费更多 token 和时间。选择方法时应在同一任务协议下并列报告准确率、延迟、输出长度和资源条件;书中测得的提升及耗时属于特定模型版本和设备,不是推理模型的固定倍率。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 评估推理模型
原书出处 · 1 处
- · 查看原书依据
min for the base model to 7 min for the reasoning model on a Mac Mini with M4 chip on the 10-sample subset, and from 13.3 min to 185.4 min on an H100 on the 500-sample dataset), which highlights one of the trade-offs of using reasoning models. Please note that these numbers were obtained in PyTorch 2.8 and can differ in different versions of PyTorch.
参考来源
- 位置:《Build a Reasoning Model (From Scratch)》第03章《评估推理模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-3.html)
- 结论依据:摘录中关于准确率与资源成本联合报告的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:min for the base model to 7 min for the reasoning model on a Mac Mini with M4 chip on the 10-sample subset, and from 13.3 min to 185.4 min on an H100 on the 500-sample dataset), which highlights one of the trade-offs of using reasoning models. Please note that these numbers were obtained in PyTorch 2.8 and can differ in different versions of PyTorch.