Request access to these research checkpoints

These are research artefacts released to support a peer-reviewed evaluation study. They are not clinical tools and must not be used to inform patient care. Access is granted to researchers who confirm the intended use below.

Pentabrid 27B: fine-tuned derivatives of Qwen3.6-27B for medical QA evaluation

Research checkpoints accompanying a controlled evaluation of how the choice of fine-tuning configuration affects medical question-answering performance at a fixed model scale.

Not for clinical use. These are evaluation artefacts. They have not been validated for any clinical purpose and must not be used to inform patient care.

Folder	Manuscript label	Configuration	Form	MedXpertQA (2,450)	Median reasoning tokens
`V13`	Configuration A	Short reasoning traces, elevated weight	LoRA adapter	26.90%	~700
`V14_27B_merged`	Configuration B	Long reasoning traces, 74% rejection-sampled + 26% curated	Merged model	41.67%	~3,700
`V15_lora`	Configuration C	As B, 50/50 data mixture	LoRA adapter	41.35%	~3,740
`V16_a64_lora`	Configuration D	As B, LoRA alpha = 64	LoRA adapter	42.20%	~4,000
`mcr_base`	Configuration E	Short expert-written traces, standard weight	LoRA adapter	41.67%	~2,435

The parent checkpoint, Qwen/Qwen3.6-27B, scores 43.76% and is not redistributed here. It is a vendor post-trained checkpoint, not an untuned base model.

V14_27B_merged is a full merged model. All other folders are LoRA adapters and must be applied to the parent checkpoint.

Summary of findings

All derivatives share one parent checkpoint and were evaluated on the identical 2,450-item MedXpertQA-Text test set with deterministic decoding and a single exact-match scorer.

Accuracy spanned 16.86 percentage points across the parent and its derivatives, and 15.30 points among the derivatives themselves.
No derivative exceeded the parent on the primary endpoint.
Four derivatives finished within 2.41 points of the parent yet produced a different outcome on 15.1% to 18.2% of items. Aggregate accuracy understates how much individual behaviour changed.
Retention of parent-correct answers ranged from 47.2% to 80.3%.
The same pattern reproduces on MedQA and MedMCQA, which were not used during configuration development.

Interpretation limits. The configurations differ in training-data content, trace length and training weight simultaneously. The study compares whole configurations and does not isolate the effect of any single training variable. Generated rationale length did not explain the accuracy differences. One training run was performed per configuration, so run-to-run variability is not captured. The analysis is exploratory: the primary benchmark partly informed the design of later configurations.

Reproducibility

Code, evaluation harnesses, the exact scoring script, per-item predictions for every checkpoint, and the checkpoint mapping are openly available without access restrictions:

https://huggingface.co/datasets/Clinical-Reasoning-Hub/pentabrid-reproducibility (DOI 10.57967/hf/9501)

Every accuracy, retention, acquisition and turnover figure reported in the manuscript can be recomputed from the per-item predictions without downloading these weights.

Citation

Model DOI: 10.57967/hf/9500

@article{agha2026turnover,
  author = {Agha, Adnan and Anwar, Eram},
  title  = {Small aggregate accuracy changes conceal substantial item-level
            correctness turnover after fine-tuning in medical question-answering},
  year   = {2026},
  note   = {Manuscript under review}
}

Authors and licence

Dr. Adnan Agha (ORCID 0000-0002-2704-8931) and Dr Eram Anwar (ORCID 0009-0006-9335-9208), College of Medicine and Health Sciences, United Arab Emirates University.

Released under CC BY-NC-ND 4.0. Access is gated so that intended research use can be recorded; requests from researchers are granted.

Updated September 2026 to align with the peer-reviewed analysis. An earlier version of this card described the fine-tuning procedure as a first-order determinant of medical reasoning and referred to a commentary submitted elsewhere. That causal interpretation was not supported by item-level analysis and has been withdrawn.

Downloads last month: -

Model tree for Clinical-Reasoning-Hub/pentabrid-27b

Base model

Qwen/Qwen3.6-27B

Adapter

(554)

this model

Clinical-Reasoning-Hub
/

pentabrid-27b