BAAI

/

bge-m3-unsupervised

like 15

Follow

Beijing Academy of Artificial Intelligence 552

Hi,

I am doing some experiments with the BGE-M3 family of models to test the impacts of unsupervised pre-training. Here are some results (R@100 on MIRACL):

It is obvious that the third row BGE Unsupervised is kind of an outlier here, because the unsupervised pre-training done on your side seem worse than 60M datapoints training on my side. I wonder if you uploaded the wrong checkpoint or that I am not using/evaluating this checkpoint correctly.

Thanks.