Development of an Eye Care Chatbot Based on Llama 3.2 1B Using 4-bit QLoRA Technique

Teguh Rijanandi (Telkom University)
Eko Risdianto (University of Bengkulu)

Journal on Informatics Visualization and Social Computing
Vol. 2 No. 1 (2026): Journal on Informatics Visualization and Social Computing (JIVSC)

Background: The ratio of ophthalmologists to the population in Indonesia is extremely low, leading to unequal access to eye care and making it difficult for the public to obtain early information regarding eye symptoms.Aims: This study aims to develop and validate an efficient, domain-specific eye care chatbot using the Llama 3.2 1B Instruct model, focusing on an optimal fine-tuning pipeline for limited hardware constraints.Methods: Utilizing an Experimental Research (Model Development) design, a QLoRA 4-bit quantization technique was applied on a Google Colab T4 GPU (15.64GB VRAM). The training dataset comprised 16,742 samples from Kaggle Eye Care and MedQuad Indonesian Translation, with a subset of 5,000 samples used for experiments.Result: After training for 3 epochs (750 steps), the Training Loss decreased from 1.3394 to 0.7188 (46.3% improvement), while Perplexity reached 2.3620, categorized as excellent. The model maintained stability with a final gap of 0.1407 between training and validation loss.Conclusion: The Meta-Llama 3.2 1B Instruct model, fine-tuned with 4-bit QLoRA, is highly effective for building a domain-specific healthcare chatbot, successfully operating within an 8GB VRAM limit while maintaining high generalization capabilities.