Concerns regarding the reproduction of F1 performance

[qqpower22]

Dear authors and friends,

I have a question to consult with you. When trying to reproduce the results, I was hoping to achieve the 92.5 F1 score on the CoNLL03 dataset as reported in the paper. According to the steps described in the paper, I sampled 10,000 instances from each dataset to construct the training set for fine-tuning. My question is: is the standard procedure to first train on the PileNER dataset for 5 epochs, and then fine-tune on this mixed dataset, in order to obtain the 92.5 F1 result?

In addition, I would like to ask another question: if we train directly on a single dataset, would the performance of the span-based method be inferior to models using the BIO sequence labeling paradigm? I am a beginner in this field and would really appreciate your guidance!