openai-api - OpenAI ChatGPT (GPT-3.5) API : How do I make a fine-tuned GPT-3. 5 模型仅从微调数据中得到答案？

Question

OpenAI 现在允许我们微调 GPT-3.5 模型。我已经使用自己的数据集测试和微调了模型，但问题是微调模型随机生成答案，而不是根据我的自定义数据集正确生成答案。

有什么方法可以让模型仅根据我自己的微调数据集进行回答吗？

Answer 1

这是一种完全错误的方法(正如您已经发现的那样)。

正如官方所说OpenAI documentation :

Some common use cases where fine-tuning can improve results:

• Setting the style, tone, format, or other qualitative aspects
• Improving reliability at producing a desired output
• Correcting failures to follow complex prompts Handling many edge cases inspecific ways
• Performing a new skill or task that’s hard to articulatein a prompt

微调并不是用微调数据集中的特定答案来回答特定问题。

你需要实现的是基于嵌入的语义搜索，如官方OpenAI documentation中所述:

When should I use fine-tuning vs embeddings with retrieval?

Embeddingswith retrieval is best suited for cases when you need to have a largedatabase of documents with relevant context and information.

By default OpenAI’s models are trained to be helpful generalistassistants. Fine-tuning can be used to make a model which is narrowlyfocused, and exhibits specific ingrained behavior patterns. Retrievalstrategies can be used to make new information available to a model byproviding it with relevant context before generating its response.Retrieval strategies are not an alternative to fine-tuning and can infact be complementary to it.

您有两个选择:

1. 自定义解决方案(请参阅我的 past answer )。
2. 使用LlamaIndex或LangChain .