Fine-tuning tips

You must care for and feed the fine-tuned set to improve training quality (based on the metrics or experience with certain test cases). Here is a summary of OpenAI’s suggestions for fine-tuning:

1. Review existing examples for issues: You might have introduced style, logic, or grammar issues into the dataset, including examples with errors. Review the material against how the model performed before and after adding the data. You can use the epoch checkpoints as a tool.
2. Gather more examples to fill the gaps: Additional training examples might show the model how to address gaps in its abilities. It is always hard to say how much is too much.
3. Include examples with errors: Sometimes, it is best to learn from the master. Let’s ask ChatGPT about including mistakes in fine-tuned examples:

   Should fine-tuning examples
   include intentional errors
   that might be expected
   from real customers?
   Yes, it's beneficial to include
   intentional errors in fine...