Benchmark of small language models for plain-language simplification in Spanish clinical texts
TL;DR
This work presents the first benchmark of small language models for Spanish clinical plain-language adaptation and introduces MEDICLARO, a corpus specifically designed for this task, and establishes a solid foundation for integrating small language models into Spanish clinical workflows.
Abstract
Medical text simplification is important for improving health literacy and making clinical information easier for patients to understand. While clinical text simplification has been widely studied in English, Spanish remains underexplored, especially for systematic adaptation into plain-language in clinical settings. This work presents the first benchmark of small language models for Spanish clinical plain-language adaptation and introduces MEDICLARO, a corpus specifically designed for this task. MEDICLARO consists of 50 clinical notes, each with three human-written simplifications produced by cognitive accessibility experts in accordance with ISO 24495-1:2023. Four families of state-of-the-art language models were evaluated through fine-tuning and prompt-based strategies. The evaluation covers simplification, semantic similarity, factual consistency, readability, and environmental impact, and is complemented by human evaluation and qualitative error analysis. The results show that Llama-3.2-3B provides the most balance profile across efficiency, robustness, and overall performance, while RigoChat-v2-7B stands out when output quality and readability are prioritized. Overall, this work establishes a solid foundation for integrating small language models into Spanish clinical workflows, offering a sustainable, patient-centered path toward digital health accessibility.