mixedbread-ai

/

mxbai-colbert-large-v1

like 48

According to the example, when calculating the similarity, I found that the mask token of query also calculated the corresponding similarity. Shouldn't the mask be removed?
The similarity before mask is deleted is:
tensor([28.4531, 21.5156, 27.0312, 16.5312, 17.3438, 20.5781], device='cuda:0',
dtype=torch.float16)
The similarity after removing mask is as follows:
tensor([10.8672, 7.2500, 10.1562, 4.9922, 6.0781, 6.9453], device='cuda:0',
dtype=torch.float16)