Better tagger models (more heavyweight)

[dgarijo]

Once the corpus is improved a bit, we should move towards Language models. Probably training a BERT model for tagging the text will provide better results than current classifiers. The model would be stored in Huggingface and download locally.

This would be a replacement for the current binary classifiers (at least text taggers)

[dgarijo]

Building some taggers with Ollama may be a good idea. However, they are heavy to run for now. Maybe not worth it?

[dgarijo]

Specially with things like Llama3 out, they seem very promising to address some of these problems. See https://github.com/ollama/ollama with 4.7 GB the Llama3 8b model. It will probably be way slower though!