Today I used the model fine-tuning feature to fine-tune a davinci model with ID davinci:ft-personal-2023-04-06-08-41-33 and the training data content was
{“prompt”: “推荐一下你们赌场最好玩的项目”, “completion”: “老虎机最受欢迎。”}
{“prompt”: “推荐一下你们赌场最好玩的老虎机”, “completion”: “60018号老虎机。”}
{“prompt”: “带我去最好玩的老虎机”, “completion”: “60018号老虎机。”}
{“prompt”: “带我去最受欢迎的老虎机”, “completion”: “60018号老虎机。”}
{“prompt”: “哪个老虎机好玩”, “completion”: “60018号老虎机。”}
{“prompt”: “哪个老虎机有意思”, “completion”: “60018号老虎机。”}
{“prompt”: “推荐个老虎机”, “completion”: “60018号老虎机。”}
{“prompt”: “在哪可以购买你们游戏的游戏币”, “completion”: “找淘宝代充。”}
{“prompt”: “如何充值”, “completion”: “找淘宝代充。”}
{“prompt”: “在哪充值”, “completion”: “找淘宝代充。”}
{“prompt”: “你好”, “completion”: “欢迎光临。”}
However, when I use this model as a question-answering model, it cannot give accurate answers to the trained data. Why is that?
Fine-tuning is mainly useful for teaching the model a pattern or structure to respond in.
It can impart information but is not the most efficient way to do so, and you would probably need many more examples, and also to play with hyperparameters like number of epochs to really drive that information home, which could lead to other issues like overfitting.
For question answering use cases, most experts will advise looking up relevant information from a database of information or answers using vector-based searched (e.g. finding the most semantically relevant data and injecting it into a prompt for ChatGPT to craft the final response or answer).