Fine-Tuning Supervisionado de LLMs para Redação Técnica em ESG: Um Estudo de Caso com Llama-3 e QLoRA

The drafting of sustainability reports (ESG) requires technical rigor and adherence to specific standards, challenges that limit the effectiveness of generic Large Language Models (LLMs). While Retrieval-Augmented Generation (RAG) approaches solve factual hallucinations, they often fail to reproduce corporate stylistic consistency. This paper investigates the use of Supervised Fine-Tuning (SFT) to specialize a Llama-3-8B model in technical writing for governance and sustainability. The methodology used the QLoRA technique for efficient tuning on low-cost hardware, employing a dataset curated from public reference reports. Results demonstrated rapid training loss convergence and superior capability of the fine-tuned model in using normative terminologies (such as Susep regulations and GRI standards) compared to base models, validating the technical feasibility of the approach for automating sensitive corporate documents.