Trivalley Meetup - ComplianceAware Finetuning

Fine-tuning makes a model better at the task. It can also make it worse at the rules.This session shows how standard supervised fine-tuning quietly dismantles the safetybehaviour a model learned in pre-training, then walks through Compliance-Aware Fine-Tuning(CAFT), which makes compliance a hard constraint inside the training objective rather than afilter on top of it. Closes with a live notebook that fine-tunes a model and measures thecompliance drift as it happens.