ShortStories.lol

This is basically an experiment to encourage me to learn how ML / AI actually works. I generated these pretty rubbish models on my laptop(!) with 7b- & 13b-parameter Llama2 models.

• Ignore the awful stories & worse ratings. I need to fine-tune some models.
• Server stats generated from nginx logs: https://www.shortstories.lol/static/report.html
• From a Flask-perspective, I'm kinda proud of how we filter categories via query parameters.
• I've recently been playing around with Google Colab and Kaggle and HuggingFace (GPU access is useful)

This repo is roughly split in two main bits:

1. A way to generate & rate short stories
2. A web app (Python / Flask) to display the stories

Interesting files include:

• generate_stories.py
• rate_stories.py
• scrape_reddit.py
• colab-generate-story.py - that runs on Google Colab and connects to a database.

Made with the following:

• Flask
• Sqlalchemy
• HuggingFace
• Langchain

Hosting

Hosted on a digital ocean droplet (which is really nice)

TODOs:

• [-] IN PROGRESS - learn more about models (& fine-tune!)
  • Fine-tune model to determine whether text is actually a story or not (eg. huggingface fine-trained model)
  • Fine-tune model to determine whether story is good [Update: trained model (on reddit data) is useless. Need better data / better fine-tuning]
• improve prometheus metrics!
• allow human votes
• use own model?!
• log in / bot detection?
• Grafana???
• think of awesome monetisation ideas
• SSL (surprisingly easy with certbot)
• prometheus metrics or something
• scrape reddit short stories / two sentence horror
• tags
• analytics!
• sort by AI votes
• get migrations working properly
• make the website look half decent
• stories page
• stories page with lots of search-param filters. eg. tones, themes, votes

Dev tips

flask db check
flask db migrate -m 'message'
flask db upgrade
flask db history

Migrations

alembic revision --autogenerate -m "init; create stories tables"
alembic upgrade head