llama-3.2-claramed

Llama-3.2-CLARA-MeD

Llama-3.2-CLARA-MeD rewrites Spanish clinical sentences in simpler Spanish for readers without medical training. It was fine-tuned on CLARA-MeD sentence pairs using LoRA. This card describes pguerrero-igutierrez/llama-3.2-claramed, the merged checkpoint used by the project's Llama demo.

This is a research model. Its outputs have not been clinically validated and require review before use in patient-facing material.

Model Details

Model Description

• Developed by: Paula Guerrero Castelló and Iker Gutierrez Fandiño.
• Model type: Autoregressive language model, fine-tuned for sentence-level clinical text simplification; LoRA weights merged into the base model.
• Language: Spanish (es) for this task.
• Base model: meta-llama/Llama-3.2-1B-Instruct.
• License: The underlying Llama weights are governed by the Llama 3.2 Community License and its associated acceptable use policy. The metadata identifies that upstream license; this card does not introduce separate terms for the authors' contributions.

Uses

Direct Use

Research and educational exploration of Spanish clinical text simplification. The model can produce draft rewrites of individual clinical sentences for review by health professionals or medical communication specialists.

Downstream Use

It may serve as a component in an assisted authoring tool or as a comparison system in simplification research. Patient-facing deployment requires separate validation of factual fidelity, readability, and suitability for the intended readers.

Out-of-Scope Use

Diagnosis, treatment recommendations, prescribing, emergency triage, or autonomous clinical decisions. The model has not been evaluated for full medical records, other languages, or unrestricted medical question answering.

Bias, Risks, and Limitations

• Simplification can omit qualifications, alter negation, numbers, or clinical relationships, or add unsupported explanations.
• Training data comes from clinical trial protocols. Performance may differ on discharge summaries, patient messages, other specialties, or regional varieties of Spanish.
• The test set contains 120 sentences. No demographic, specialty, or literacy-level breakdown is reported.
• The split is random at sentence level. The notebook does not demonstrate document-level separation or a cross-split duplicate audit.
• Automatic text metrics do not establish medical correctness or patient comprehension. No clinical validation or patient comprehension study is reported.
• Biases from the base model and training corpus may remain. Sensitive patient information should not be submitted to a public demo.