2 devlogs
2h 25m 37s
This project uses a distilledGPT2 encoding and a generic Transformer-based architecture, trained on a portion of a Huggingface database (corbt/all-recipes), to train a 83 M parameter generative model. Use it to find your next great recipe (1% of the time) or your next disaster (99% of the time… hey I don’t have any H100s). Disclaimer: you shouldn’t probably take these recipes seriously unless you’re absolutely sure it’s fine.
