A LoRA fine-tuned adapter for Qwen/Qwen3-14B on the MedMCQA dataset — 182K medical multiple-choice questions covering 21 subjects from Indian medical entrance exams (AIIMS/PG style).

Model Details

Developed by: James Oon (@jamezoon), SUTD MSTR-DAIE Deep Learning Project
Model type: Causal LM with LoRA adapter (PEFT)
Base model: Qwen/Qwen3-14B (dense, 14B parameters, BF16, standard transformer)
Language: English
License: Follows base model license (Qwen3)
Adapter size: ~81 MB (adapter_model.safetensors)

Intended Use

Medical multiple-choice question answering. Given a clinical question and 4 options (A–D), the model selects the correct answer with a step-by-step explanation. Subjects covered include Physiology, Anatomy, Biochemistry, Pathology, Pharmacology, Surgery, Medicine, Dental, Gynaecology, Paediatrics, and more.

Not intended for real clinical decision-making. This is a research/educational model.

How to Get Started

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

base_model_id = "Qwen/Qwen3-14B"
adapter_id = "jamezoon/qwen3-14b-medmcqa-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter_id)

messages = [
    {"role": "system", "content": "You are a helpful tutor for pre-med students preparing for medical entrance exams. Answer the following multiple choice question by thinking step by step, then give the answer."},
    {"role": "user", "content": (
        "Question: Which of the following is the most common cause of mitral stenosis?\n"
        "Options: A. Rheumatic fever  B. Congenital  C. Infective endocarditis  D. SLE\n"
        "Think step by step. Then respond in the format:\n"
        "Explanation: ...\nAnswer: <one of A, B, C, D>"
    )},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Training Details

Dataset

MedMCQA — 182,822 training samples, 4,183 validation samples
21 medical subjects (Dental, Surgery, Medicine, Pathology, Pharmacology, etc.)
Each sample: question + 4 options + correct answer (1-indexed) + explanation
Formatted as chat messages with system/user/assistant roles

Training Procedure

Hyperparameter	Value
Training steps	1,000 (max_steps — ~8.7% of 1 full epoch)
Epochs	1 (partial)
Per-device batch size	4
Gradient accumulation	4 (effective batch = 16)
Learning rate	2e-4
LR scheduler	Cosine
Warmup steps	100
Max sequence length	512 tokens
Precision	BF16
Optimizer	AdamW (fused)
Max grad norm	1.0

LoRA Configuration

Parameter	Value
Rank (r)	16
Alpha (α)	32
Dropout	0.05
Target modules	q_proj, k_proj, v_proj, o_proj
Trainable parameters	20,971,520 (0.1418% of 14B)
Bias	none

Hardware & Training Time

Hardware: NVIDIA GB10 Grace Blackwell (NVIDIA DGX Spark), 121 GB unified CPU+GPU memory
Training duration: ~26 hours total (1,000 training steps + 5 evaluation passes × ~95 min each)
Actual training steps: ~83 minutes (1,000 steps at ~5s/step)
Framework: PyTorch 2.x, HuggingFace Transformers, PEFT 0.18.1, TRL (SFTTrainer)

Architecture Note

Qwen/Qwen3-14B is a standard dense transformer (not the Qwen3.5 hybrid variant). It does not use GatedDeltaNet linear attention layers, making it fully compatible with standard CUDA training without special kernel requirements.

Evaluation

Training Loss Progression

Step	Train Loss	Token Accuracy
10	2.972	55.5%
50	1.502	69.5%
100	1.124	76.6%
200	1.061	77.3%
600	1.052	77.3%
1,000	1.068	76.7%

Validation Loss (Dev Set, 4,183 samples)

Checkpoint	Eval Loss	Token Accuracy
Step 200	0.9825	78.96%
Step 600	0.9746	79.02%
Step 800	0.9681	79.14%
Step 1000	0.9664	79.18%
Best (saved)	0.9649	79.20%

Eval loss improved consistently throughout training, indicating good generalisation.

MCQ Accuracy Comparison (Dev Split, 4,183 samples)

Model	Accuracy	Notes
Qwen3-14B zero-shot	27.4%	Format failures common (~12.5% None responses)
Qwen3-14B + LoRA (this adapter)	TBD	Evaluation in progress

Zero-shot accuracy is low primarily due to format non-compliance — the base model frequently fails to output a clean A/B/C/D answer in zero-shot settings. LoRA fine-tuning addresses both format adherence and domain knowledge.

Per-Subject Zero-Shot Baseline (for reference)

Best subjects: Anaesthesia (38.2%), Psychiatry (37.5%), Radiology (31.9%) Weakest subjects: Orthopaedics (10.0%), Skin (11.8%), Anatomy (19.2%)

Comparison with Qwen3.5-9B Adapter

	Qwen3.5-9B adapter	This adapter (Qwen3-14B)
Base model params	9B	14B
Architecture	Hybrid (GatedDeltaNet)	Standard transformer
Trainable params	7.1M (0.079%)	21.0M (0.142%)
Best eval loss	0.9669	0.9649
Best token acc	78.7%	79.20%
Adapter size	28MB	81MB

Citation

bibtex

@inproceedings{pmlr-v174-pal22a,
  title     = {MedMCQA: A Large-scale Multi-Subject Multi-Choice Dataset for Medical domain Question Answering},
  author    = {Pal, Ankit and Umapathi, Logesh Kumar and Sankarasubbu, Malaikannan},
  booktitle = {Proceedings of the Conference on Health, Inference, and Learning},
  year      = {2022},
  publisher = {PMLR}
}

Framework Versions

PEFT 0.18.1
Transformers (latest as of March 2026)
TRL (SFTTrainer)
PyTorch 2.x + CUDA