开源了 GPT 的问题测试集一共 39000，方便大家冷启动批处理问问题

[newlin]

Our project provides 500 carefully designed questions that can be used to test and evaluate the performance of Generative Pretrained Transformer (GPT) models. The questions cover a variety of topics and domains and are designed to comprehensively examine the performance of GPT models Comprehension, creativity and logical reasoning skills.

https://github.com/zgimszhd61/benchmark4GPT

[cryptogems]

大致看了一下，这些问题拿到对应的答案之后评价指标是什么，and 中文问题，是不是某种程度上其实没法发挥模型的真实性能，毕竟隔着一层语言的 layer （个人想法）