Can I please ask if Base version or Instruct version is right choice for fine tuning LLM model?
For example, I use the LLM model for multiple text classification. Should I choose Llama3.1 8b or Llama3.1 8b Instruct? And can this be generalized to other models? Like for Gemma2 9b vs Gemma2 9b Instruct, Mistral 7b vs Mistral 7b Instruct, Mistral-Nemo-Base-2407 vs Mistral-Nemo-Instruct-2407 and so on…It’s not clear to me from the documentation…
TL;DR Use the instructional version of the models.
As Always - It depends; Mostly on your resources and expectations.
In most cases, ppl who don’t require domain adaptation or significant differences in alignment - will want to work with the Instruct version, for LLaMA or any other LLM.
For example,
- Prompting isn’t a thing on the base version, and you’ll need to use few-shot and other techniques to get the model to understand what you want.
- Using Instruct versions allowing you to achieve almost anything with a prompt, but you get stuck in this “ping-pong UX”.
In your case, for multiple-text classification,
-
Create a test set.
-
Start with prompting a model.
-
Do the same with different models.
-
Fine-tune the best-performing model. Test the FT model. Good luck.
Sahar Whether you fine-tune a base model or an instruct model depends on the behavior you want your final model to have and the structure of your data.
Sometimes, it makes more sense to start with a base model rather than further specializing a model that has already undergone instruction tuning. Other times, starting from an instruct model is better if you want to preserve its instruction-following behavior while adapting it to your specific task.