Precise training data of Combo

[djstrong]

In the readme:

Combo Pre-trained continued from original BERT on 2017, 2018, 2019 SEC 10K dataset

but in the paper:

train a Combo Model on top of the last checkpoint of BERT-Base Uncased. This training was done in parallel with FinBERT Prime, using SEC2019 for the first 250,000 and using SEC1999 for the last 250,000.

What is correct?