Hello,
I have finetuned GPT-3.5-turbo-0125 to sound like a particular robot personality using curated data from the robot’s creators. This data covers simple questions from preferences to abilities (i.e., the robot can’t walk or move objects). However, when attempting to use the fine-tuned model, it cannot/will not generate proper nouns, such as names or places, even if you give the name or place in the prompt. Does anyone know what the cause of this may be or how to fix it? Thanks.
Heya! Welcome to the dev community!
Do you have a sample of your fine-tuning dataset? ie how did you train?
Do you have an example of prompts… or what you’re trying to achieve?
It is logical that if you don’t have any fine-tune coverage of the inputs that would generate or repeat names, that inference of the underlying “chat” abilities would be lower.
If you want it to say Detroit, Microsoft, or Jim’s Robot Factory, you’ll need chat about the topics that might invoke factual city or company responses.
You can try one of the checkpoints and see if that can perform your task without being overtrained on it.