This repo contains a nootebok which shows in details how to finetune roberta (as the example I used the polish version: sdadas/polish-roberta-base-v2) model for part-of-speech tagging task. It uses modern API from Huggingface ecosystem and is heavily based on https://huggingface.co/docs/transformers/tasks/token_classification.
I used this template to train two taggers for polish language: wkaminski/polish-roberta-base-v2-pos-tagging and wkaminski/polish-roberta-base-v2-cposes-tagging.
Dataset used in training: ipipan/nkjp1m