Evaluation Variability with Random Instruction Selection

[leeesangwon]

First of all, thank you for sharing your excellent research with the community.

I have a question regarding the code implementation for evaluations using the M-BEIR dataset. It appears that the instruction is chosen randomly when conducting evaluations. This leads me to believe that if the random seed is not fixed, the evaluation performance could vary depending on which instruction is selected. Do you have any experimental results where you fixed a specific instruction for evaluation?