Results on BEIR

[hfawaz]

Hello thanks for this!

Based on the ColbertV2 paper: https://arxiv.org/pdf/2112.01488

In table 5, you have very low results on NQ for example compared to some models in the leaderboard for retrieval: https://huggingface.co/spaces/mteb/leaderboard

Is there a reason behind this?

If single-vector retrievers from mteb leaderboard are better at most datasets, why use the multi-vector ?

Thanks