Entrenamiento de un modelo de lenguaje de gran tamaño para el análisis de casos jurídicos desde el enfoque positivista del derecho

Abstract

Abstract: This thesis addresses the complete development cycle of a large-scale language model that enables the automatic analysis of legal cases from the perspective of judicial reasoning, limited to criminal judgments issued by first-instance judges of the Ecuadorian Judiciary, from a positivist approach to law. The current Constitution of the Republic of Ecuador establishes the constitutional guarantee of reasoned decisions in Article 76, paragraph 7, subparagraph l), which requires that all resolutions issued by jurisdictional bodies contain a verifiable logical and normative justification. To implement this principle, the legal argumentation methodology developed by Manuel Atienza was adopted, which breaks down judicial reasoning into seven structural components. The Mistral-7B-Instruct-v0.3 model was trained using the QLoRA fine-tuning technique, generating a dataset of 152 synthetically annotated criminal judgments, complemented by a RAG system with supplementary documentation that yielded 14,338 normative fragments. The creation of the AI Judicial Auditor prototype was deployed in HuggingFace Spaces, achieving an average Atienza F1-Score of 0.928, with a synthetic expert evaluation of 92.2/100. This evaluation was generated using large commercial AI models, Gemini 3 Pro and ChatGPT 5.1, resulting in a high level of confidence in the assessments performed.