Additional vectors usage in training process to increase model accuracy

[notiqq]

Hi,

I'm looking for potential ways to increase trained model accuracy and was thinking that the increasing amount of vectors might help.
During the model training process for Ukrainian language I have noticed that spaCy has downloaded package with data (probably I would assume it was vectors package for Ukrainian language??) But in documentation, it says that spaCy works on the base of tok2vec vectors that are actually generated in pretraining process. And it actually means that there is no direct way to substitute existing vectors package with something else?
Could you please clarify whether it's pretrained actually or it uses some additional tokens that being downloaded?
And maybe there will be another suggestions on potential ways to increase model accuracy?
P.S. for training process the standard UD package for Ukrainian language has been used.

[polm]

The way vectors are used in spaCy is that pretrained vectors are used as input when training a tok2vec model, which is then used as the input to other pipeline components like the tagger, parser, and NER. The pretrained vectors are also exposed in the final pipeline and can be used in similarity calculations, for example.

For the pretrained models, the small models do not contain static vectors, but medium and large ones do. All models have a trained tok2vec layer.

You can add existing vectors to a model, but you need to convert them first with spacy init vectors. You can read more about how to specify custom vectors for training models here, or how the static vectors are used here.

And maybe there will be another suggestions on potential ways to increase model accuracy?

Generally I would suggest using the quickstart with the "accuracy" option to establish a baseline, and then clarify what you specifically want to improve the accuracy of.