turk-rad-llm-lumbar-mini-v2

Early research preview for Turkish lumbar MRI report structuring

Overview

turk-rad-llm-lumbar-mini-v2 is an early research-preview LoRA adapter exploring structured information extraction from Turkish lumbar MRI reports.

The model investigates whether Turkish free-text radiology reports can be transformed into a structured output format using parameter-efficient fine-tuning. This public repository is intended as a limited proof-of-concept and does not include the full clinical preprocessing workflow, annotation prompts, post-processing logic, validation dataset, or deployment pipeline.

Disclaimer

This model is not approved for clinical use. It is intended for research and educational purposes only. Outputs must be reviewed by qualified medical professionals and must not be used as the sole basis for diagnosis, triage, treatment, or patient-facing decisions.

Model Details

Property	Value
Developed by	Ahmet Harun Turgan, MD
Base model	google/medgemma-4b-it
Method	Parameter-efficient fine-tuning with LoRA
Language	Turkish
Domain	Lumbar spine MRI reports
Output type	Structured text / JSON-style output
Status	Research preview
Clinical use	Not approved

Data

The adapter was trained on a limited internal research dataset of anonymized Turkish lumbar MRI-style reports.

Dataset governance, annotation strategy, quality-control procedures, and formal validation methodology are being documented separately for academic review. No source clinical reports, full annotation prompts, preprocessing scripts, or validation pipeline components are included in this public preview.

Intended Use

This model may be used for:

• Research on Turkish clinical NLP
• Educational demonstrations
• Method development for structured radiology report extraction
• Non-clinical benchmarking under appropriate license terms

This model must not be used for:

• Primary clinical diagnosis
• Automated clinical decision-making
• Patient-facing applications
• Clinical deployment without independent validation, institutional approval, and applicable regulatory authorization

Usage

Basic loading is supported through PEFT-compatible workflows. Users are responsible for adapting inference code, validation logic, and output parsing to their own research environment.

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("google/medgemma-4b-it")
model = PeftModel.from_pretrained(
    base_model,
    "ahmet-harun/turk-rad-llm-lumbar-mini-v2"
)

This public preview does not provide the full production inference wrapper, prompt strategy, post-processing logic, or schema-validation pipeline.

Evaluation

A small qualitative sanity check was performed during development. The results should not be interpreted as rigorous clinical benchmarking.

Formal evaluation, including field-level metrics, multi-reader assessment, and external validation, is outside the scope of this public preview and is planned separately.

Limitations

This is an early proof-of-concept model. Important limitations include:

• Domain restriction to lumbar MRI-style reports
• Limited research-preview dataset
• No external validation
• No prospective clinical validation
• Possible malformed structured output
• Possible incomplete extraction
• Possible terminology errors
• Possible inconsistent field assignment
• Not approved for clinical deployment

All outputs require expert human review.

Reproducibility and Access

This repository provides high-level model information for transparency and research visibility.

The following components are not publicly released:

• Source clinical reports
• Full annotation prompts
• Data preprocessing scripts
• Post-processing and JSON-repair logic
• Validation dataset
• Deployment pipeline
• Clinical workflow integration components

Access to additional materials may be considered only through formal academic collaboration and after appropriate institutional, ethical, and regulatory review.

License and Safety Notice

This model is a LoRA adapter derived from google/medgemma-4b-it and is subject to the applicable Gemma / MedGemma / Health AI Developer Foundations terms of use.

Users are responsible for compliance with all relevant license terms, local regulations, clinical validation requirements, institutional policies, and health regulatory requirements.

Citation

If you reference this research-preview model, please cite:

@misc{turgan2026turkradllm,
  author       = {Turgan, Ahmet Harun},
  title        = {turk-rad-llm-lumbar-mini-v2: Early Research Preview for Turkish Lumbar MRI Report Structuring},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {https://huggingface.co/ahmet-harun/turk-rad-llm-lumbar-mini-v2}
}

Please also cite the relevant MedGemma source model according to its original model card and license requirements.

Author

Ahmet Harun Turgan, MD Radiology Resident · Istanbul, Türkiye

For academic collaboration or validation discussions, please contact the author through Hugging Face.