medical-subject-headings-01075e1f·1 events·first seen Aliases: Medical Subject Headings
Researchers evaluate five small open-source LLMs (up to 9B parameters) on identifying semantic relationships between biomedical concepts, introducing MeSH-Rel-4K, a 4,000-relationship dataset derived from Medical Subject Headings. Three adaptation strategies are compared: standard prompting, Chain-of-Thought prompting, and fine-tuning. Fine-tuning yields a 34.1 percentage point average F1-score improvement, demonstrating that targeted fine-tuning can overcome reasoning limitations of parameter-constrained models for specialized domain tasks.