Small LLMs for Biomedical Claim Verification: Cost-Effective Fine-Tuning, Structural Dataset Shortcuts, and Cross-Domain Generalization

In three linesThree small LLMs (Phi-3-mini 3.8B, Qwen2.5-3B, Mistral-7B) fine-tuned via QLoRA for biomedical claim verification. Mistral-7B outperforms GPT-4o and GPT-5 (+12% F1) on 1,008 training examples. Study identifies structural artifact in SciFact and demonstrates robust cross-domain generalization.

Summary generated by Claude — human-verified