Small error in the sentence transformer example
#9
by fbilhaut - opened
The comment in the sentence transformer example says:
Sentence Transformers calls Softmax over the outputs by default, hence the scores are in [0, 1] range.
After checking the numbers and the code, it seems that these scores do not come from a softmax, which isn't applied by default, but from the default activation function which is a sigmoid.
See method predict() in CrossEncoder.py .
Confirmed — verified against the current sentence_transformers/cross_encoder/model.py source: get_default_activation_fn() returns nn.Sigmoid() when num_labels=1 (this model's setup), not Softmax. Opened a PR to fix the wording: #16