The application of Large Language Models in the medical domain in Vietnam presents significant challenges due to the scarcity of domain-specific training data. Traditional Supervised Fine-Tuning methods are often ineffective on small datasets, leading to overfitting and a failure to capture the complex nuances of medical language. To address this, our study proposes Token Value-Aware Fine-Tuning, an advanced fine-tuning method, as an alternative to SFT on low-resource datasets. The TVAFT process uses the Gemini API to generate judgments, followed by computing prominent weights to guide the pre-trained model’s focus toward critical tokens during training. To evaluate TVAFT, we fine-tune the PhoGPT-4B-Chat model for a medical question-answering task using a dataset comprising 250 Vietnamese medical question–answer pairs. Experimental results demonstrate that TVAFT significantly improves evaluation metrics compared to SFT, including BLEU, ROUGE, BERTScore, and token-level F1. These findings establish TVAFT as an effective approach for tasks requiring deep understanding under limited data conditions, highlighting its potential for developing specialized Vietnamese health consultation systems.