For DPO I only had around 700 preference examples, so not much data at all. That took about 12 minutes to train on a single GPU.

Pretraining was obviously a a lot slower, the 125M model took roughly half a day.