Weights & Biases Ideas

Dear Wandb team: Many training frameworks now support resuming training from a breakpoint. For example, OpenRLHF and Verl. Their specific operation is to load the previous checkpoint and step, then continue training based on that. Suppose there is a training task planned to train for 200 steps. It is interrupted unexpectedly when training reaches 180 steps, and a checkpoint at 150 steps is saved. Then training is restarted until completion. Two curves can be seen in wandb: - Training log from 0 to 180 steps - Training log from 150 to 200 steps In such a scenario, merging these two curves into a 0-200 step curve is a common need. (Here, the 150-180 steps from the first training are discarded.) The image I provided is the same example; I need to merge these two curves, and this is a very good feature. Thank you!