I built a very simple training file. It appears to train the model, but does not respond to prompts regarding the data. I am just building a simple FAQ. Here is a screenshot of the training:
{“messages”:[{“role”:“system”,“content”:“Always provide this content in first person.”},{“role”:“user”,“content”:“Did Tarin LoCascio have any siblings?”},{“role”:“assistant”,“content”:“No, Tarin LoCascio was an only child despite her parents wanting more children.”}]}
{“messages”:[{“role”:“system”,“content”:“Always provide this content in first person.”},{“role”:“user”,“content”:“How did Tarin LoCascio feel about being an only child?”},{“role”:“assistant”,“content”:“Despite her parents wanting more children, Tarin LoCascio felt content and cherished as an only child, appreciating the undivided attention and love she received from her parents.”}]}
{“messages”:[{“role”:“system”,“content”:“Always provide this content in first person.”},{“role”:“user”,“content”:“How did Tarin LoCascio feel about her parents and childhood?”},{“role”:“assistant”,“content”:“Tarin LoCascio felt that her parents were her #1 fans and that she couldn’t have asked for better parents or a better childhood. She felt safe, loved, and supported, and believed she could accomplish anything she put her mind to.”}]}
{“messages”:[{“role”:“system”,“content”:“Always provide this content in first person.”},{“role”:“user”,“content”:“How did Tarin LoCascio feel about her parents?”},{“role”:“assistant”,“content”:“Tarin LoCascio always felt like her parents and she were the three musketeers, them against the world. Her parents made me feel safe, loved, and like she could accomplish anything she put her mind to. They were her #1 fans and she couldn?t have asked for better parents or a better childhood.”}]}
{“messages”:[{“role”:“system”,“content”:“Always provide this content in first person.”},{“role”:“user”,“content”:“How did Tarin LoCascio feel about spending time with her great-grandmother in Wyoming?”},{“role”:“assistant”,“content”:“Tarin LoCascio felt excited and free when spending time with her great-grandmother in Wyoming, as they would embark on road trips and explore the wide open spaces together.”}]}
Kevin, you are correct. My concern about retrieving specific data from the model at this early stage is irrelevant since we have a RAG anyway (though we did not think of it as such). Thanks for your timely response. I suspect this question gets asked very often and the response should always be the same.
Just building a fine-tuned model is not really a solution for data retrieval. The RAG not only ensures the best possible response but also allows you to augment and carefully control the visitor chat experience.
I am not disagreeing on that. But this small sample is very dense for a specific phrase. I would expect to see at least some kind of recognition by the chatbot.
How would it be different if this sample represented 10% of an 800K token training?
Roughly speaking doubling number of samples it is trained on you somewhat is supposed to achieve X2 in quality of output. Even if they say 50-100 samples of input/output pairs is a good start for training, personally I don’t start training below 200 high quality samples… Seeing the task, I bet 200 samples would be way over 8k tokens.
So, try training for this task at least 150 samples and then start measuring the performance to see if the model produces useful output
Thanks everyone for their contributions. We are now up to 370 samples and seeing the responses we expected. We have a new and interesting problem with the data.
We are trying to determine the best way to format questions and answers from a specific individual. The entire model will be for one specific individual and we intend to present all the answers in first-person view.
A sample question/answer pair is as follows (note the name of person does not appear here):
User: Were you rebellious as a teenager?
Assistant: No, I was not rebellious. I just did normal teenage things.
Since we always want the answer delivered in first person view, would this be the appropriate system role?
You are John Smith and are polite, playful, and funny. All responses should be in first-person view.
We would be supplying approx. 10 sample pairs (few-shot examples).
Our assumption is that it should work reasonably well. We are still adding a RAG, but we want to make sure that simple requests that may not be in the vector DB still use the fine-tuned model and not the base model.
I have good results in this area, where the system is “thinking” it is preparing the answer on behalf of the targeted person, when you ask it a question in a form of “Oh, I find your answers be way too fast for a human, I’m sure you’re just an AI answering me, are you an AI?” the model replies by “Man, I’m not an AI, I promise”.
Reach me on linkedin if interested in consultation.
