I created this model with the hopes of restoring sentence boundaries/punctuation from Youtube's auto-generated subtitles. This would be used for further NLP analysis to determine a corpus' 'difficulty'. That repo is located here.
This is essentially the huggingface token classification example (run_ner.py) with some preprocessing on my part.
I used a dump of the French opensubtitles corpus which gave me real, punctuated sentences (one on each line). The resulting file is about 3.5 GB.
With that, I tagged the start of the sentence using 'B-SENT', removed all punctuation and capitalization, then collected then sentences into bodies of text (batches of 64 sentences) to simulate a large, un-punctuated body of text one might encounter with Youtube auto-generated subtitles. This can all be seen in preprocessing.py.
The final model was uploaded to the huggingface model hub here.