What are the parameters used for the published performance (from F1=87.1 in the dev set of Squad1.1)?
#2
by
marcusborela
- opened
I'm trying to repeat the published performance (from F1=87.1 on the dev set of Squad1.1), but I'm not getting such a good result (I got F1=85.7). Would it be possible to know which parameters were used for me to adjust here? (doc_stride, handle_impossible_answer, max_answer_length, etc)