llama31-medqa-qlora

library_name: transformers tags:

• llama
• medical
• usmle
• multiple-choice
• text-classification
• causal-lm
• qlora
• peft
• instruction-tuned

---

Model Card for david125tran/llama31-medqa-qlora

This model is a QLoRA-fine-tuned version of Meta-Llama-3.1-8B-Instruct on the MedQA USMLE 4-option multiple-choice dataset. It is trained to answer USMLE-style medical questions by selecting one of four answer choices: A, B, C, or D.

---

Model Details

Model Description

• Model type: Causal language model (decoder-only) used as a multiple-choice classifier
• Base model: meta-llama/Llama-3.1-8B-Instruct
• Fine-tuning method: QLoRA via PEFT / SFTTrainer (TRL)
• Language: English
• Task: Medical multiple-choice question answering (USMLE-style)
• Author: David Tran (david125tran)
• License: Inherits the license and use restrictions of meta-llama/Llama-3.1-8B-Instruct and the MedQA dataset.
• Finetuned from model: meta-llama/Llama-3.1-8B-Instruct

This model is intended as a research and learning artifact in my “LLM-Fine-Tuning-Lab” repo, demonstrating how an instruction-tuned base model can be adapted into a domain-specific multiple-choice classifier using modern fine-tuning techniques (QLoRA, LoRA, 4-bit quantization).

Model Sources

• Training code / experiments: (GitHub) https://github.com/david125tran/LLM-Fine-Tuning-Lab
• Dataset: GBaker/MedQA-USMLE-4-options on the Hugging Face Hub

---

Uses

Direct Use

The model can be used to:

• Answer USMLE-style medical multiple-choice questions with 4 options (A–D).
• Serve as a research baseline or teaching example for:
  • QLoRA fine-tuning
  • PEFT/LoRA adapters on top of an instruction-tuned base model
  • Logits-based multiple-choice scoring

Typical usage pattern:

1. Format the question + options into a single prompt.
2. Feed it through the model in inference mode.
3. Compare the final-token logits for the tokens "A", "B", "C", "D" and choose argmax.

Downstream Use

Downstream, the model may be:

• Integrated into offline analysis pipelines for medical education research (e.g., question difficulty analysis, distractor quality, or model-assisted item generation).
• Used as a starting point for further fine-tuning on:
  • Specific medical subdomains (cardiology, neurology, etc.)
  • Different exam formats or MCQ banks
  • Chain-of-thought or rationale-style training

Out-of-Scope Use

This model must not be used for:

• Real-time clinical decision making or patient care.
• Providing medical advice to patients or the public.
• Any safety-critical application where incorrect outputs could cause harm.
• Automation of high-stakes licensing or certification decisions.

It is a research/educational artifact and should be treated as such.

---

Bias, Risks, and Limitations

• The model is trained entirely on MedQA (USMLE-style) questions, which are exam-style, synthetic, and biased toward specific curricula and style of reasoning.
• It may:
  • Overfit to exam-style phrasing and perform worse on free-form clinical narratives.
  • Reflect dataset biases in disease prevalence, demographics, and clinical practice norms.
  • Produce confident but incorrect answers without uncertainty estimates.

The model has only been evaluated on a small subset (~200 examples) of the MedQA distribution during this experiment and has not been rigorously validated for robustness, calibration, or fairness.

Recommendations

• Treat outputs as unverified suggestions, not as ground truth.
• Do not deploy in clinical settings.
• If used in research, combine with:
  • Human expert review
  • Calibration analysis (e.g., confidence vs. accuracy)
  • Error analysis, especially on high-confidence wrong answers

---