-
Notifications
You must be signed in to change notification settings - Fork 1.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Integrate sentence transformers into benchmarks #843
Conversation
I checked the data and it seems ok. No answers as well as long answers are removed from NQ dev set. Lets revert the config and fix the mypy bug, then we are ready to merge from my side. What do you think @brandenchan ? |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This all looks good to me. Just one very tiny comment. Its ready for merge as far as I'm concerned once the mypy bug is fixed and the config is reverted.
Hey @brandenchan I fixed mypy and reverted the config. Could you double check, also with the conflicting docs files, so that we can merge? |
Ok nice! The conflicting docs seem to be because new arguments were added to functions in master, we updated doc strings and the api documentation was regenerated. I would in each conflict case take the change from master |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Once documentation conflicts are resolved, this PR is ready for merge
Would be nice to compare the new sentencetransformers models, especially the
SentenceTransformer('nq-distilbert-base-v1')
and how it compares against DPR finetuned on NQ.Made sure to use cosine similarity (so had to use ES doc store).
Documents for Sentence Transformers should be a list of [title, text] lists.
Preliminary results do not look better than DPR:
for 100k docs we have mAP of 82.7 SBert vs 86.5 DPR
Details: