Raphael
Results
3
comments of
Raphael
you can do ``` import service_streamer service_streamer.service_streamer.WORKER_TIMEOUT = 120 ``` before importing the Streamer
yes, pretty sure the position_ids are incorrect during training with left padding. You would need to write a custom DataCollator and "manually" create position_ids for each instance in your batch
minimal example: https://gist.github.com/raphael-sch/c38c4dbaecce62566bc3f07559678360