About
Goal:
- Play with machine learning and web technologies
- Eventually, contribute to the field of Craptology (see Journal of Craptology)
- Have fun
Techstack:
- Corpus build from IACR eprint abstracts from 1996 to end of 2021
- GPT 345 fine-tuned on Google Colab using https://github.com/minimaxir/gpt-2-simple with 5000 steps
- Server running on AWS EC2
- Static HTML and Javascript served by FastAPI/Uvicorn
- Completion by gpt2tc
- Communication to gpt workers via FastAPI websocket
Author: Thomas Pöppelmann
Todos:
- Allow to set k and p values (k is currently at 900 so that gpt considers the title)
- Game to allow users to play "fake or real paper"
- Add model for other domain than crypto (system security?)
- Provide more models (e.g., larger gpt models) or options (differently trained)
- Extend corpus beyond eprint (e.g., older LNCS conference abstracts not put on eprint)
Impressum
-
Angaben gemäß § 5 TMG:
-
Thomas Pöppelmann,
Wilhelm-Riehl Str. 14,
80687 Munich
Kontakt:
-
E-Mail: ThomasPoeppelmann@googlemail.com