the problem was docker, that had as entrypoint a bash script with python train.py --epochs=300 hardcoded, so I guess it was never acutally running the task setup from clearml.