DAFT: Building a Medical Report Interpretation System with 1.1B-Parameter TinyLLaMA, Hallucination Rate Only 2.1%

The medical AI field faces a dilemma: large models have high deployment costs and high hallucination rates, while small models lack professional capabilities. The DAFT project uses domain-adaptive fine-tuning and hybrid architecture design, leveraging the 1.1B-parameter TinyLLaMA model to achieve 97.9% accuracy and 2.1% hallucination rate in blood test report interpretation tasks, significantly outperforming medical domain baseline models like BioBERT and ClinicalBERT, providing a feasible solution for lightweight medical AI applications.

Blood test reports are important basis for medical diagnosis, but over 60% of patients cannot understand them accurately, leading to anxiety and increased communication costs between doctors and patients. Traditional manual interpretation is inefficient, and general-purpose large models have hallucination risks due to the extremely high accuracy requirements of medical scenarios, making them difficult to deploy directly.

DAFT adopts a hybrid architecture: the deterministic component achieves 100% accurate numerical extraction through regular expressions and rule engines; the generative component converts structured data into patient-friendly explanations based on the fine-tuned TinyLLaMA model. The layered design ensures both medical rigor and humanized expression.

The 1.1B-parameter TinyLLaMA model is selected, and LoRA fine-tuning (r=16, α=32) is used to inject domain capabilities. The training data consists of 850 manually labeled samples (split 8:1:1), cross-validated by 3 medical experts (κ=0.83). Performance saturates when the number of samples exceeds 500, showing high data efficiency.

The system supports PDF/image uploads. The process includes OCR recognition, numerical parsing, anomaly detection, intelligent interpretation, and result presentation, taking approximately 2.3 seconds in total. The tech stack includes React+TypeScript for the frontend, FastAPI for the backend, and the model is deployed on Hugging Face Spaces.

Evaluated by 5 medical experts via a triple-blind protocol, DAFT has a hallucination rate of 2.1% and an accuracy rate of 97.9%, significantly outperforming BioBERT (9.4% hallucination rate), ClinicalBERT (11.8%), and BioGPT (7.1%). In robustness tests, the accuracy across lab formats ranges from 87.5% to 100%, and it maintains 94.7% accuracy even with 5% OCR errors.

DAFT can be trained and deployed on consumer-grade hardware (12GB VRAM GPU), promoting the inclusiveness of medical AI. The system clearly prompts that it cannot replace professional medical advice, reflecting ethical responsibility, and is positioned as an educational auxiliary tool to improve patients' health literacy.

DAFT's open-source code includes training scripts, model weights, and deployment guidelines, and the related paper was published at the ICCET international conference. In the future, it can be extended to more types of medical reports, support multiple languages, and integrate personalized recommendations, providing a model for small model applications in vertical domains.