---
license: cc-by-nc-nd-4.0
language:
- en
base_model:
- Qwen/Qwen3.6-27B
library_name: peft
tags:
- medical
- clinical-reasoning
- medxpertqa
- lora
---
# Pentabrid 27B — Medical Reasoning Models

Models supporting the Nature Medicine Matters Arising response to Oermann and Vishwanath (2026), "General-purpose large language models outperform specialized clinical AI tools on medical benchmarks" (s41591-026-04431-5).

## Core finding
At fixed scale (27B) and offline (no retrieval), the fine-tuning procedure is a first-order determinant of medical reasoning accuracy. A short chain-of-thought fine-tune collapsed the base model's reasoning length and dropped MedXpertQA accuracy by 16.9 points. A corrected long chain-of-thought approach restored reasoning length and recovered accuracy to near parity with the base model. The claim is recovery to near-parity, reversing the regression, not superiority over the base model.

## Contents of this repository
| Folder | Model | Manuscript label | MedXpertQA (2,450) | Median reasoning tokens |
|---|---|---|---|---|
| V14_27B_merged | Long chain-of-thought (primary) | Fine-tune B | 41.7% | 3,700 |
| V15_lora | 50/50 data-mix ablation (adapter) | — | 41.4% | 3,738 |
| V16_a64_lora | LoRA alpha=64 ablation (adapter) | — | 42.2% | ~4,000 |

For reference, the untuned base (Qwen3.6-27B) scores 43.8% (3,100 tokens), and the short chain-of-thought fine-tune (Fine-tune A / V13) scores 26.9% (~760 tokens). V15 and V16 are two independent null ablations confirming the recovery is robust to configuration.

Manuscript-to-repository mapping: Fine-tune A corresponds to V13 and Fine-tune B corresponds to V14.

## Usage
V14_27B_merged is a full merged model. The V15 and V16 folders are LoRA adapters and must be applied to the base model Qwen3.6-27B.

## Reproducibility
Evaluation scripts, the exact scoring script, and the full results table are in the companion dataset repository: https://huggingface.co/datasets/Clinical-Reasoning-Hub/pentabrid-reproducibility

## Authors and license
Dr. Adnan Agha (ORCID 0000-0002-2704-8931) and Dr. Eram Anwar (ORCID 0009-0006-9335-9208), UAEU College of Medicine and Health Sciences / Tawam Hospital. IP: UAEU Application #2442. License: CC-BY-NC-ND-4.0.