Code for the bert on NQ baseline and TAP method used in the paper

[abhi1nandy2]

Could you share code for the BERT on NQ baseline and TAP method used in the paper?

[cvittorio]

The TAP submission to the internal baseline was from a different group.
That code was based on https://github.com/IBM/translucent-answer-prediction
but it was apparently modified for the submission to the TechQA leaderboard and the modified code has not been open-sourced.

The system for the BERT on NQ baseline was described here: https://arxiv.org/pdf/1909.05286.pdf
and is not open-sourced.
The reason why we released this code is because it outperforms the BERT on NQ and we could open-source the code.
The metrics are:
F1 = BEST_F1 = 56.62
HA_F1@1 = 34.08
HA_F1@5 = 42.58