English models' Accuracy Evaluation values

[ojo4f3]

Hello!

I have a question regarding the English model stats displayed on the spaCy website: https://spacy.io/models/en. The Accuracy Evaluation tables for the en_core_web_sm, en_core_web_md, and en_core_web_lg models are exactly the same (although the en_core_web_lg model's Named entities (recall) value is 0.86 instead of 0.85 like the others). The en_core_web_trf model's values are unique.

I understand these values may be correct but wanted to make sure since I'd assume that the larger models would have better results.

Thanks,

Colton

[adrianeboyd]

English has enough training data that larger models and vectors don't make a huge difference. All the raw evaluation numbers are here (the website tables are dynamically directly generated from this data):

https://github.com/explosion/spacy-models/tree/master/meta