Hi, i’ve successfuly fine-tuned the unsloth/Phi-3.5-mini-instruct model with some simple data, using their free Colab notebook.
Here’s my data:
Can apples can go in the red bin?,yes
Can apples can go in the blue bin?,no
Can pears can go in the red bin?,yes
Can pears can go in the bluebin?,no
Can grapes can go in the red bin?,yes
Can grapes can go in the blue bin?,no
Can plums can go in the red bin?,no
Can plums can go in the blue bin?,yes
Once training is complete I wish to be able to test the model with the following:
list all items that can go in the red bin
Or
should i put pears go in the red bin or the blue bin?
My question for this forum is, am I expecting too much? As it just doesn’t work. Should this work using Unsloth etc. How deeply is my data integrated within the model?
Any assistance in aiding my understanding is greatly appreciated.
After more research i’ve discovered that supplying context with my training data would help, e.g:
Context: apples can go in the red bin. pears can go in the red bin. grapes can go in the red bin. plums can go in the blue bin.
Question: Can apples can go in the red bin?
Answer: Yes.
My problem however now is that the Colab (Google Colab) doesn’t use this kind of template.
I was assuming that to provide context there would be an API call/overload etc. The reality, of course, is ridiculously simple, you just supply the context in the body of the query, the following is a typical RAG template query e.g:
"Answer the question based on the context below. If you can't answer the question, reply, "I don't know". Context: apples can go in the red bin. pears can go in the red bin. grapes can go in the red bin. plums can go in the blue bin. Question: list everything that goes in the red bin."