This model ranked 1st place in open llm leader board, However this model has lower performance in supervised fine tuning.

#8
by andreaKIM - opened

Hello, Thanks for giving such a great model with instruction tuning.
This model showed highest benchmark in the open llm leaderboard however when i try fine tune with SFTtrainer, this models shows lower performance with comparison to just llama2-13b-chat model.
I did even try to tune hyperparameters for optimization several times, it didn't work.
Is there any specific reason for this reason?

Hello,

Are you using this model (upstage/SOLAR-10.7B-Instruct-v1.0) for finetuning? If so, you might not achieve the highest score.
You should instead use our pretrained model (upstage/SOLAR-10.7B-v1.0) for finetuning.

(If you have already been using our pretrained model for your finetuning process, please disregard the above.)

I recommend the following settings:

  • learning rate: 1e-06 ~ 2e-06
  • at least 5 epochs
Limerobot changed discussion status to closed

@Limerobot Thanks for recipes for fine tuning
Is there any reason fine tuned model get poor performance when they fine tuned once again?
I am just curious if there exist some specific reason for this situation.

Sign up or log in to comment