I got the same error message when attempting to use the same .csv format as for training. After a little trial and error, what eventually worked was supplying the data in the following format, as e.g. used for Vision tasks in CreateML: test_data/ ├─ label_0/ │ ├─ 0.txt │ ├─ 1.txt │ ├─ ... ├─ label_1/ │ ├─ 0.txt │ ├─ 1.txt │ ├─ 2.txt │ ├─ ... Here, the files 0.txt, ... containing only the text strings. Hope this helps you!