Preprint
Jul 2026
Every Sample Counts: Supervised Fine-Tuning of Language Models with Pointwise Constraints
This work proposes a novel alignment framework that addresses the gap between per-sample constraints and average loss by enforcing per-sample constraints while still minimizing an average loss, and develops an augmented Lagrangian approach tailored to this formulation.
Ignacio Hounie, Ignacio Boero, A. Ribeiro
· 2 citations