23.4 OOB ML Models Have Longer Pipeline Training Times

Why do the 23.4 OOB ML Models Have Longer Pipeline Training Times?

Issue: "Significantly" longer training times may be observed when training the newer 23 versions of UiPath's OOB ML models in comparison with the older 22.10 and prior models.

Example - 11 hours total training time on v22.10.2 Vs 100+ hours on v23.4.1 (both running on GPU & using the same dataset)

Explanation:

Via internal communications with the Product team, a longer training time is expected for the 23 versions of the model vs previous models prior to 23.

The trade-off for using the newer models, is that the new versions of the model should have a higher accuracy for extraction.

  • If faster training time is preferred over higher accuracy, the recommendation is to use the 22.10 versions of the models instead of the 23 versions of the models for now.

Note: An enhancement to improve training times is planned for future releases of the models.