Bridging the Medical Knowledge Gap: 4-bit QLoRA Fine-Tuning of Llama-3-8B for Biomedical Reasoning
Large language models (LLMs) show promise for biomedical reasoning, but the compute needed to adapt them keeps this work out of reach for many researchers. This study examines whether Meta Llama-3-8B can be adapted for biomedical reasoning on a single NVIDIA T4 GPU (16 GB VRAM) using 4-bit NormalFloat (NF4) quantization and quantized low-rank adaptation (QLoRA). The adapted model, named Bio-Llama-3-Expert, was fine-tuned on the Medical-O1-Reasoning-SFT dataset with approximately 42 million trainable parameters (0.52% of the model) and evaluated on two small, study-level sets. On a 50-question PubMedQA subset (YES/NO/MAYBE), it reached 72.0% accuracy (36/50; 95% CI, 58.3–82.5%), compared with a 42.0% majority-class rate in that subset and a reported 45% for untuned Llama-3-8B with naive prompting. On a separate 69-question reasoning-complexity set, overall accuracy was 79.7% (55/69) and declined from 92.0% on direct-recall questions to 65.0% on multi-hop questions. The MAYBE class was the weakest (3 of 9 correct). The study is limited by small evaluation sets, short training runs, baseline comparisons that are not paired, and the absence of clinical validation. The results are therefore preliminary evidence of computational feasibility, not of clinical readiness.
Authors
- Aditya Verma
Publication Details
- Journal
- Pittsburgh Undergraduate Review
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5195/pur.2026.166
- Primary Topic
- Artificial Intelligence in Healthcare and Education
- Type
- article
- Field-Weighted Citation Impact
- 0.00