hyperopt score is -1 if run fails
see main.py the following part:
if training_finished: hyperopt_score, n_epochs = get_hyperopt_score(cfg, mlflow_run) else: hyperopt_score, n_epochs = -1, cfg["n_epochs"]
This is a problem if the hyperopt score is negative (e.g. when hyperopting for negative rewards). Try np.nan instead of -1.