Inspiration
Can a language model recognize when its own certainty is unjustified? And can it maintain a correct answer when challenged with false information?
These questions inspired Knowing That You Do Not Know, a red-teaming project developed for Hack Apertus Track 1A.
What it does
The project investigates five reproducible reliability failures in Apertus v1.5 70B, involving arithmetic verification, contradictory answers, stated confidence, self-correction, and responses to unfounded user challenges.
How we built it
We developed five Python experiments using the hosted Apertus inference API, with reproducible execution through Docker and GNU Make.
The experimental ideas and design are by Giuseppe Faraone. Python implementation and English documentation were prepared with ChatGPT assistance.
Challenges and lessons learned
The greatest challenge was distinguishing observed failures from their possible causes. Additional exploratory controls helped identify alternative explanations and refine our conclusions.
The project reinforced an important lesson: a reliable AI system should not merely produce correct answers, but should also know when its confidence is justified and when a correction is warranted.
Log in or sign up for Devpost to join the conversation.