I fine-tuned a model last night with the following dataset (one of the examples is given below)
The fine tuning process is all running good and success.
{“messages”: [{“role”: “system”, “content”: “You are a helpful assistant.”}, {“role”: “user”, “content”: “Who is Anita Fung?”}, {“role”: “assistant”, “content”: “Anita Fung is a scientist with a master’s degree in clinical nutrition, used to work in three hospitals as dietitian in Australia .”}]}
However, when i try to query the new model with question: “Who is Anita Fung?”, the new model answered me with strange information unrelated to the above fine tune dataset.
How can I solve this issue? It looks like the new model doesn’t learn things I taught.
Basically GPT is a massive autocomplete machine, and fine tuning will bias it to be more likely to select certain words over others.
Fine tuning it doesn’t let it learn new facts. It’s more suited to training it to respond a certain way. But you still have to give it the facts and context of the question.
For example, it’s suited to making responses more sarcastic or more suited to customer service. You can use fine tuning to train a personality or a style of writing into the responses. It’s very suited for classifications too, like if you have a list of traffic violations and want to put it in the appropriate category.
If you want to give it facts, you might want to look into the “chat with PDF” tools, or build something similar yourself with embeddings.