Which checkpoint gets evaluated?


Does the server-side setup evaluate only the last checkpoint model, as the following parameters from the starter kit suggest?

checkpoint_freq: 100
checkpoint_at_end: True
keep_checkpoints_num: 5

Or are you evaluating the best one? If so, at what frequency do you save checkpoints during evaluation?


Hello @andrey_kolobov

We use the latest checkpoint during the rollouts.

1 Like