Lumina: A Fine-Tuned LLaMA 3.1 Model for Empathetic Psychological Support

Abstract

Mental health remains a pressing global concern, with the World Health Organization (WHO) estimating that nearly 1 in 8 individuals worldwide suffer from a mental disorder, and the vast majority do not receive adequate care due to stigma, limited awareness, and a shortage of mental health professionals. In response to these challenges, this paper presents Lumina, a fine-tuned conversational AI assistant based on the LLaMA 3.1 framework, designed to simulate empathetic, human-like psychological support. Lumina leverages advanced language modeling techniques-including parameter-efficient fine-tuning with LoRA and 8-bit quantization-to provide a confidential, accessible, and judgment-free environment for users to share their emotions and receive guidance. The system emphasizes natural language understanding, emotionally sensitive interaction, and a userfriendly design, operating through a Flutter-based front-end and a scalable, Modal-hosted backend. By enabling contextaware and emotionally attuned dialogue through fine-tuning, and evaluating performance using empathy-centric dialogue metrics such as emotional alignment and linguistic matching, Lumina bridges critical gaps in mental healthcare delivery, especially for underserved populations. This paper outlines the system’s architecture, including dataset curation, fine-tuning methodology, emotion-aware interaction design, and efficacy evaluation. Lumina represents a cost-effective, scalable solution for democratizing mental health support and reducing societal stigma around seeking help.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 1 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.