Difference between model-best and model-last

[umayerr]

I took a spaCy blank model and trained that for ner with chemical data. At the end of the training process I found two different models (model-best and model-last) as output. I am wondering the difference between these two models? Which model should I use for entity recognition from the text? If I want to continue incremental training which model would be the perfect one to start?
Note: I am using spaCy v3.

[adrianeboyd]

Here's a closely related discussion: #9848 (comment)