unexpected response even it is Fine tuned with custom dataset
It giving the response that could not really understand the input data.
It giving the response that could not really understand the input data.
Hi coder could you share a bit more about your datasets and how you fine tuned the model?
i used the unsloth notebook to train the mistral models ! <<
so you can use the alpaca format : or change it to suit your needs , in this case i owuld train the model on some instruct ( alpaca ) datasets : then when outputting the model to GGUF to make sure you first train the model with a Q/A prompt : As the gguf GPT4all and software uses this :
BUT:
If your going to use it as a openAI type model ( you need to train it with the chat ml format before you output it ) this model was not trained with the ChatML template so if you want to use tools or funciton calling then : train using the chatml format and gguf it like this :
If you want to dump a lot of knowledge into the model then you would use the text completion method : ( then it will generate text well ) ---
NOTE: because you trained it in any of these mode: you just need to choose your method to output the model at the end in gguf :
so After you have trained your model : then you can edit your tokenizer.config and remove an underlaying templates : then train for alignment for a few step to implement the correct template for youor desired usage : ( then it will imprint this on to the model ) :
then when you output the model and host it local or on server it will respond as expected !
Really for software its a Q/A format and for servers its ChatML
SO probably make the same model and output both variations :
this strategy allows you to train for anything ! then align the model to your desired prompt and output template .. the save as gguf :
the best qwhich has worked the cheepest and well on colab(free) was the unsloth ( to run local you need python 3.10( windows 50/50 ) - Linnx 100% ( all need cuda setup working )
Runs fine i the free colab ! >>> Unsloth also keep updating the notebooks they share ! << so as they change the server setups and update the bitsandbytes and peft etc it moves with them so if you have problems just get a new notebook as they are maintaied !>>>
its easy to miss a stage !! Alignment is important !<<<<<< KEY !
It giving the response that could not really understand the input data.
Hi coder could you share a bit more about your datasets and how you fine tuned the model?
My dataset is in the form of
##Instruction:
##Response:
And for Fine tuning model, I uses peft, lora