It seems worse than the old version
#2
by davidbnu - opened
Many thanks for sharing this valuable model. I’ve tested it, but it appears to perform less accurately than the earlier version (dh-unibe/trocr-kurrent-XVI-XVII). Do you have any insights into what might be causing this?
Thanks a lot for the feedback! The two trocr-kurrent models are actually trained for different historical periods:
dh-unibe/trocr-kurrent is optimized for 19th-century documents, while dh-unibe/trocr-kurrent-XVI-XVII is intended for sources from the 16th to 18th centuries. Have a look at the model cards, where the training material is listed.
We might consider renaming them to make that distinction clearer.