> trained from scratch on 80B tokens of historical data How can this thing possibly be even remotely coherent with just fine tuning amounts of data used for pr

> trained from scratch on 80B tokens of historical data How can this thing possibly be even remotely coherent with just fine tuning amounts of data used for pretraining?