Why would you need to customize GPT models?
GPT models are already very useful without any customizations. When your user types a request, you, as a programmer, could simply forward the request to the GPT model (such as GPT-3.5 or GPT-4), and, in many cases, the unaltered response from the model is good enough. However, in many cases, the responses aren’t good enough. There are three categories of problems with responses:
- Non-text functionality: In some cases, the response you want is not text-based. For example, you may want to allow your user to turn a light on or off, perform complex math, or insert records into a database.
- Lack of context: Models can’t accurately answer questions if they haven’t been exposed to the data that contains the answer. Despite being trained with immense amounts of data, there’s a lot of data that LLMs haven’t been exposed to. At the time of writing, the cut-off date for data used to train GPT 3.5 and GPT...
