Comment by pests

I was curious, they train a 1900 base model, then fine tune to the exact year:

"To keep training expenses down, we train one checkpoint on data up to 1900, then continuously pretrain further checkpoints on 20B tokens of data 1900-${cutoff}$. "