MD-Reasoner-8B
This repository describes the 8B MD-Reasoner model introduced in our paper, Enhanced Medical Diagnostic Reasoning in Small Language Models Using Reinforcement Learning. The model artifacts will be released once dataset-sharing approval is finalized and a compliant distribution pathway is in place.
Availability
This Hugging Face repository intentionally does not include:
- model weights
- LoRA adapter weights
- training data
- tokenizer or checkpoint files
The underlying research model was trained on data derived from MIMIC-IV. Because MIMIC-IV is distributed under credentialed-access terms through PhysioNet, adapter release on Hugging Face is currently pending dataset-sharing approval.
Until that process is complete, this page lets the work be cited, discovered, and reviewed without redistributing restricted artifacts.
Model Summary
MD-Reasoner-8B is the 8B-parameter variant in the paper. It applies:
- A cold-start supervised fine-tuning stage on curated clinical diagnosis examples.
- A reinforcement learning stage using Dr.GRPO / GRPO-style optimization with rewards for diagnostic correctness and output structure.
Base model:
Release Status
The research uses training data derived from MIMIC-IV, a credentialed-access clinical dataset distributed via PhysioNet. Current PhysioNet guidance indicates that derived datasets or models should be shared under the same agreement as the source data. We are therefore holding artifact upload until the appropriate approval and release path are in place.
Intended Use
This repository is intended for:
- reading the model summary
- citing the paper
- reviewing benchmark results
- understanding the training approach and release status
It is not intended for:
- model download
- production inference
- clinical deployment
- autonomous diagnosis
Results
Reported results from the paper on the held-out MIMIC-IV evaluation subset:
| Model | Setting | Diagnostic accuracy |
|---|
| Qwen3-8B-Base | Zero-shot | 49.54% |
| Qwen3-8B-Base | Zero-shot CoT | 50.77% |
| MD-Reasoner-8B | SFT + RL (GRPO) | 66.56% |
Training Data
The paper reports high-level dataset construction as:
- 11 target diagnosis groups
- 3,667 processed samples after filtering
- 790 curated cold-start SFT samples
- 2,554 RL post-training samples
- 323 held-out evaluation cases
The underlying source is MIMIC-IV. This repository does not include any patient data or note-level examples.
Risks And Limitations
- This page does not provide runnable weights.
- The research model is not a medical device.
- The study covers 11 diagnosis categories and text-only inputs; performance outside that scope is unknown.
- Clinical outputs require expert oversight and must not be used for autonomous decision-making.
Paper
- Title: Enhanced Medical Diagnostic Reasoning in Small Language Models Using Reinforcement Learning
- DOI:
10.1016/j.ijcce.2026.02.002
Citation
1@article{gebreab2026mdreasoner,
2 title = {Enhanced Medical Diagnostic Reasoning in Small Language Models Using Reinforcement Learning},
3 author = {Gebreab, Senay and Musamih, Ahmad and Salah, Khaled and Jayaraman, Raja},
4 journal = {International Journal of Cognitive Computing in Engineering},
5 year = {2026},
6 doi = {10.1016/j.ijcce.2026.02.002}
7}
Contact
