One point to consider for the example you brought up about specified GPT model variants for certain areas. All the model was trained on is being able to "understand" linguistics and languages. It is a common mistake by people that they think that GPT 3 or Chat GPT know everything they saw during training from wikipedia etc. This is not the case at all. You just need all that data to teach a LLM how a language works, the syntax, the spelling, how to express yourself etc. To do so, you need to have all kinds of word combinations occuring multiple times. The model itself does not remeber the factual information it received uring trainng in the first place. It just learns how different words / tokens depend on each other when they appear in a specific sequence. This makes it look like the model knows these facts, whereas in reality it just predicts the next most likely token.

So for the example you brought up, the model is already perfectly fine. Where you need to finetune it is for specific kind of languagaes, like medicine, legal, etc.

But in general, never rely on the factual outcome of such a model, it does not know it and it cannot know it by design. It just happens to output it correctly because based on stochastics the outputted words had the highest probability based on your input promt.