Sub-Billion, Super-Frontier: Small Language Models Rival Zero-Shot Frontier LLMs on General and Literary Relation Extraction

The study evaluates small language models (SLMs) ranging from 360M to 3B parameters for relation extraction (RE) across general and literary domains, demonstrating that the Qwen2.5-0.5B model fine-tuned on general-domain data achieves a micro-F1 score of 0.83, outperforming zero-shot models like GPT-5.4 and Claude Sonnet 4.6. This highlights that with targeted task adaptation, compact models can effectively compete with larger frontier models, offering a more resource-efficient and privacy-sensitive alternative for practitioners, particularly in settings where deployment constraints exist. The findings suggest that task-specific data can significantly enhance performance, making SLMs viable for practical applications in relation extraction tasks.