SMALL LANGUAGE MODELS

OPPORTUNITIES FOR COLLECTIVE HEALTH IN THE AMAZON

Abstract

The Brazilian Amazon Region exhibits profound connectivity inequalities that, in its isolated areas, preclude the use of cloud-based artificial intelligence tools and stable telemedicine networks. However, health professionals and community health workers operating in these territories carry personal mobile devices. This study proposes and implements an informational and technological framework based on small language models, executed entirely offline on these devices, with retrieval-augmented generation supported by official manuals from the Brazilian Unified Health System. The framework articulates documentary curation organized into six thematic pillars within a compact edge database; the production of a synthetic dataset comprising one thousand clinical dialogues in Brazilia Portuguese, incorporating regional terminology and 18,5% refusal examples; the quantized fine-tuning of the base model; and an evaluation protocol centered on information quality. The evaluation, conducted with an unseen set of 33 items, recorded a dosage-hallucination rate of 3%, understood as the occurrence of numeric dosages absent from the retrieved context. The results preliminarily indicate that, in the experiments conducted, the curation, organization, and indexing of information was more determinant of system reliability than model scale, suggesting a path to mitigate informational asymmetries without dependence on global satellite infrastructures.