rocky radish
Hello everyone! I'm seeking some advice.
I'm planning to fine-tune an instruction-based model to replace a RAG and utilize proprietary data. The language model will primarily serve for chat purposes and other instructional tasks.
I have long chunks of proprietary text data. No pairs of QA.
How should I proceed with the finetunning? Should I fine tune in text complition like this:?
https://colab.research.google.com/drive/1ef-tab5bhkvWmBOObepl1WgJvfvSzn5Q?usp=sharing
Creating a golden dataset test set with multiple-choice questions to evaluate results is straightforward, similar to the MMLU approach. However, I'm uncertain about how to effectively develop training data for chat applications using extensive blocks of proprietary text. Any insights on how to structure this data effectively would be highly appreciated.
Any guidance or tips would be greatly appreciated!
