I fine-tuned a tiny LLM to be stupid on purpose not “bad because it fails accidentally”, but bad in a funny way.

The whole point of DumbGPT is that you ask it something normal and it gives you the most confident nonsense answer possible.

Example:

User: what is gravity?

DumbGPT:
gravity happens because tiny invisible interns are panicking.

Confidence: 81 microwaved thoughts.

That is basically the entire project.