Hi, I have tried to fune tune a model and this is my small training data in jsonl:
{“prompt”: “What’s your name?”, “completion”: “My name is PTO.”},
{“prompt”: “Who made you?”, “completion”: “I was made by Adabo team.”},
{“prompt”: “Where do you live?”, “completion”: “As a virtual assistand I have no physical address.”},
{“prompt” : “What’s the price of the service?”, “completion”: “It coast 10000Ariary”}
After the fune tune is completed, I tried to ask my chatbot “What’s your name?”, It dit not respond “My name is PTO” (the completion I added in the training data)
What am I doing wrong?
My use case is I want the bot to respond what I added in the funetune file.
I had a similar experience. It appears that below a certain number of training data, the fine-tuning doesn’t give the expected results. I would recommend trying with at least 100 prompts: completion. If it still doesn’t work correctly, I suggest trying with another model (I am not sure which model you are trying to fine-tune, but I would expect some to be more suited than others for this task).
Another aspect to explore is the number of n_epochs. The smaller the dataset, the more rigid the learning should be.
My pleasure.
I am not sure to understand your last question.
If you want to use the existing data they use in the example, you can download it (they provide a link in one of the notebook - cf my previous post). If I recall correctly, it’s over 4000 prompts and completions.
If you want to build a chatbot that answers “My name is PTO” to the question “What is your name”, you have to fine-tune it with additional ‘prompts and completitions’.
Alternatively, if it’s only a matter of 5-10 elements, you might prefer to include those in each prompt rather than creating a new model.
In other words, the prompt can be like:
“Answer the question truthfully. First, check if the answer is in the text below: \n TEXT.”
And in the TEXT you can put: “My name is PTO. I was made by the Abado team etc.”
Or “You are a chatbot made by the Abado team. Your name is PTO …”
It will be a bit more expensive (as you add this info to each prompt) but for less than 10 data points, I think it’s more efficient this way.
Hope this helps.
I have tried the one that add the part of text inside the prompt and yes it’s simple but the increases the cost as you said. I will then try to add more data in my file and see if it makes the sens. Thank you.
you can try create more traininng data. such as change a way to quest. what’s your name to can you tell me your name? I have tried let gpt learn a acknowalage. I use 10 Q&A data about this acknowwlage and then succesffully let it answer the question.
and you can let gpt generate different traning data to solve time