What is [initialize] vector='model' and what are the differences between stock models?

[source19069]

We have a trained a bioinformatics model using spacy for inference of numeric test values associated with specific disease treatments, with good scores. The input datasets were created using the spacy en_core_web_sm model.

We'd now like to compare our results by generating other models using different stock models as a base, so we'd like to understand the difference between, say the en_core_web_sm and en_core_web_lg models.

Obviously we can just rebuild the datasets using different as bases and train new models, but what is the effect of using

[initialize]
vector='model'

in the config file during training?

We used 'en_core_web_lg' for the existing model training and the statistics and results look good, but we would really like to know what effect this parameter actually has, if any, on the trained model.

Thank you.

[polm]

As noted in the docs, that just determines which static vectors are included in your model. Depending on your settings those can be used as input features during training. See the static vectors docs.

The difference between the small, medium, and large models is whether they include word vectors, how many they include, and the size of the model. See here for details.