Scores for the retrieved passages

[c3-pakazad]

I have a question concerning the approach for computing the scores for the retrieved docs. I have a text query with N tokens, where N <= query_maxlen. Assuming that attend_to_mask_tokens=False, do we expect the scores associated with documents retrieved for this query to be less than N or less than query_maxlen?

[okhat]

attend_to_mask_tokens=False doesn't affect this.

So, "less than query_maxlen"

[okhat]

attend_to_mask_tokens=False is about earlier layers in the transformer, not the MaxSim scoring of colbert