Rank and scores
In the description you mention that what matters is the relative order for the sentence similarity task and not absolute values of the score similarity. I have a dataset of mathematical definitions and methodologies .What puzzles me is that even when the query has nothing relevant with the dataset I get a high score so I can not trust this model 's output. Also I have to mention that is the only model I have found which does an excellent job for the Greek language. Do you know how I cam bypass the problem of questions which are out of context? Setting a threshold doesnt seems to work. Really need your opinion on that
I am not quite clear about your question, but if a query has nothing relevant with the dataset, this model should give a relatively low score compared to relevant (query, document) pairs, although the absolute score will still be above 0.7
Also, if you find the model's output is not reliable, how do you determine this model does an excellent job for the Greek language?
Apologies for the confusion. As I mentioned earlier my dataset holds mathematical definitions and corresponding questions written in the Greek language. When the query is relevant to the context of the dataset, meaning that there is an answer which corresponds to the question, then I always retrieve the correct definition using the embeddings produced by the model. So it does an excellent job in my test case. The only problem I am facing is with questions that are out of context and how I can filter them out. Fine tuning perhaps? I am sorry if this is not the right place to post that kind of question. If that's the case feel free to delete.
I have found a response of yours in the discussion of intfloat/multilingual-e5-large which answers my question. So i will close the topic