[Feature] Support model training on multiple datasets

[rasa-pa]

📄 Description

Currently, models can only be trained using a single dataset. This prevents users from combining complementary data sources or leveraging multiple datasets to improve model generalisation.

🎯 Objective

Enable users to train a single model on multiple datasets within one training session.
This should allow selecting and combining two or more compatible datasets directly in the training configuration.

[piotrgrubicki]

@rasa-pa , could you share why importing both datasets to a single project does not work for you?

[rasa-pa]

There are several reasons:

• The datasets may come from different (similar) products or different batches of the same product. We want to keep track of that.

• Different labels. Each dataset may have a different label representation. Also new labels can be introduced. We work incrementally and merged datasets become larger and harder to navigate.

• Independent dataset ownership. To speed up the labelling process it can be several annotators working with a folder each.