Hello sir, I have been following you on RU-vid. I know you are a master of AI in Chinese. I have an OCR project for Chinese characters at the temple gate or main hall. I would like to contact you for some advice on datasets, models or some tips in this regard. I come from Vietnam, so my understanding of Chinese is not good. It's hard to find meaningful information on this topic.
In brief, the autoCoT method just uses a series of questions to do zero-shot CoT, and then use the output of zero-shot CoT as prompts to do few-shot learning. The questions are selected by cluster.
不是很好理解的视频。我的理解是 有个h函数,它既要拟合大语言模型的输出 又要拟合supervision 函数的输出。根据原文,supervision function 是”In practice, the supervision functions can be as simple as keywords, regular expression, or knowledgebase querying. The details are discussed in Section 4.“ 在测试时,如果h输出的分低(不知道这个分怎么来的。我猜它输出的是符合supervision function的probability。h中训练时跟每个supervision function的期望输出算过loss,所以当大语言模型与某一个supervision function的输出相抵触时,probability可能会比较低),那么就开始依次检查查具体哪个supervision function输出低。将低的function对应的feedback prompt返回给大语言模型做纠正