Additional readme (delete before merge)

New features

In bots/gpt2tuned you can find fine-tuning folder which contains all the instruments to learn tuned gpt2 model (on labeled shortjokes.csv dataset)
In api/ folder there is a small backend (FastAPI) + Jinja2 template, after run of api.py the generated jokes will be displayed
Jokes can be evaluated by classifiers and length criteria

Problems

Fine-tuning requires a lot of resources (GPU+memory), so I've used our university JupyterHub. Guess it is better to deploy it on AWS or hugging-face
I can show how modified model generate the jokes on our JupyterHub (they look pretty logical but still offensive)
Some classifier parameters must be changed (need more time to play with them)
THERE IS NO OFFENSIVE FILTER. Unfortunately,I didn't add anyone but found suitable solution - https://developers.perspectiveapi.com/

Things to do

Due to PEP, it is better to use .toml files for builds - https://peps.python.org/pep-0518/
Solve all the problems
Classify jokes using NLP methods
Expand web interface
Maybe containerization. Docker-compose with ML part + Web