You can change it for other entities, or even at runtime – although using a very different beam width at runtime may produce worse results.
I’ve put a small fix back into the latest spaCy release that I think was helping training times for ner.batch-train previously.
The fix is to make an update through the greedy parser with some probability (say, 50% of the time). The greedy updates are faster, and are easier to learn from in the early iterations. It’s a bit of a hack though, and lacking clear evidence of whether it was generally good, I disabled it.
You could try this by doing pip install spacy==2.0.11.dev0. This will install the brand new development release I just uploaded. If you rerun your training with this version of spaCy, how does your accuracy look?