Exploring Post-Training Alignment of Small Language Models for Biomedical Data-to-Text Generation: A Case Study of Medication Leaflet
The study compared four post-training alignment methods—SFT, DPO, ORPO, and GRPO—on Qwen2.5 models for generating medication leaflets from structured data. ORPO showed the best in-domain semantic performance, while GRPO proved most robust against changes in regulatory styles across different datasets.