google/gemma-2-9b · base vs instruct model

Jul 2

I have an use case similar to question and answering on my custom data set, should I be starting with base model or instruct model for this exercise? or is there a thumb rule on when to use base and when to use instruct model?
@lysandre any guidance is greatly appreciated

Renu11

Google org Jul 11

Base model is good for general language understanding but may not follow instructions well. It needs fine-tuning for specific tasks. Instruction tuned Model is pre-trained to follow instructions, good for tasks like question answering or chatbots and it also may need less fine-tuning.
Base models can be used for tasks where more control over fine-tuning needed or if we have a large custom dataset. Instruction tuned models are used for tasks that involve following instructions and require good performance out-of-the-box.

Please refer to this similar issue answered in Kaggle for more details. Thank you

saireddy changed discussion status to closed Jul 11