[Predictive Intelligence] After transferring a trained ML solution from the source instance via an Update Set, the solution training fails multiple times with error message 'Error while training solution' on the target instance - Support and Troubleshooting

[Predictive Intelligence] After transferring a trained ML solution from the source instance via an Update Set, the solution training fails multiple times with error message 'Error while training solution' on the target instance Issue All Predictive Intelligence solutions have a Word Corpus associated with it. Although this is not mandatory for Classification solutions, it will auto-generate the Word Corpus when you first train the solution -

To transfer a trained Predictive Intelligence solution to another instance via an Update Set, in the Related Links section, click on "Add solutions to the current update set" -

This will add the ML solution artifacts, such as solution definitions, template records, and predictive model statistics to the current update set. Once the Update Set has been retrieved and committed on the target instance, the solution should provide predictions without having to retrain the transferred solution.

Cause On occasions, due to certain limitations, it may not have transferred all of the ML solution artifacts, and hence you may need to retrain the Predictive Intelligence solution on the target instance. However, if the associated Word Corpus record is also missing, the solution training will fail multiple times with the error message 'Error while training solution' on the target instance.

Note: Trained Predictive Intelligence (ML) Solutions generate records in table [ml_model_artifact] and each record contains an attachment [sys_attachment] with the trained solution, as the instance is the source of truth for all solution training. These trained solution attachments can be very large in size and Update Sets were designed to transfer configuration records to other instances, not large attachments, hence it may not capture all of the [ml_model_artifact] records for a solution. Therefore, to re-generate these missing trained solution attachments, you will need to train the Predictive Intelligence solution again on the target instance.

Resolution When you click on "Add solutions to the current update set" in the Related Links section on a solution definition on the source instance, check the Update Set to ensure it has captured the Word Corpus records too -

If these records do not exist in the Update Set, you can use the "Force to Update Set" functionality on the Word Corpus records to ensure these records are captured. If the Word Corpus definition does exist on the target instance, but it still fails to train the solution, open the system property [glide.wordvector.upgrade_time_frame] and set it from 180 (days) to 0, so that it will train the Word Corpus and generate the missing [ml_model_artifact] records and their attachments, when you submit the solution for training. Once the solution training has successfully completed and the Word Corpus has been regenerated, set the system property [glide.wordvector.upgrade_time_frame] back to 180 (days), as it only requires training once every 6 months.