Using AIGC Confidence Levels to Quickly Obtain Golden Training DataSets

Like ChatGPT, many AIGC already have powerful cognitive, interactive and conversational capabilities. Enterprises are moving towards fine-tuning training to produce better business domain-specific models. However, if companies want to conduct fine-tuning training, they need to find representative data for labeling from the large amount of data produced by daily operations. These methods may employ expert screening or trained human labor. These effective human resources are not easy to obtain, so the output of exclusive customized models is usually slow and costly. This paper proposes a method to quickly and effectively filter out high-quality datasets from camera snapshots or spectrograms generated by light and sound detection devices, reducing the time and cost required for fine-tuning training. First, define the business purpose conditions and use these defined conditions as prompt words. Secondly, the image and these prompt words are synchronously transmitted to the large language model for question-answering service. After these answers are integrated and organized, the large language model will give different scores from 0 to 9. According to the response score and the $\mathbf{2 0 / 8 0}$ rule, the pictures with scores of $\mathbf{9, 8}$, and 7 are classes as ‘positive’, and the pictures with scores of 0 and 1, two classes as negative classification. Finally, we fine-tune these images and perform inference. We found that this method has an accuracy of $97 \%$ for a test set of images manually annotated by experts. We call this method “AIGC-Confidence Levels” Method. This method can be used to automatically annotate images, thereby obtaining low-cost but high-efficiency golden data required which AI fine-tune models