The score ratio is very vague.
#9
by AoiEugeo - opened
I found that the score range is not 0 - 1. When I input two identical sentences, the score is sometimes 13 and sometimes 14. For completely unrelated sentences, the scores could be 6, 4, etc. I want to know what the maximum and minimum scores are and how to determine a high similarity.
Filter out scores below 7.5
Yes, this model doesn't use a Sigmoid or some other function to map the scores to 0...1. You can either apply your own, or use a score threshold matching what the model usually outputs, like @xldistance proposes.
- Tom Aarsen