Push-Back AI: Fine-tuning LLMs to doubt, disagree or refuse

Beschrijving

Tired of over-helpful AI? This session shows how to engineer friction as a feature: LLMs that surface uncertainty, ask first, and push back when needed. We’ll start with a quick A/B demo (including visible inner reasoning) and then walk the practical path: lightweight QLoRA fine-tuning on a reasoning-centric base, running via Ollama, and the creation and curation of synthetic datasets that explicitly train doubt, disagreement, and guardian behavior. You’ll leave with a repo, models, a micro-dataset, and a tiny eval script to try at home. This isn’t a prompting talk... it’s hands-on model adaptation for developers.

Periode	21 okt 2025
Evenementstitel	Tilburg Tech Tuesday XXL 2025
Evenementstype	Congres
Locatie	Tilburg, NederlandToon op kaart
Mate van erkenning	Regionaal