README
MedShift
Training-free domain generalization for medical VLM hallucination mitigation via retrieval-augmented decoding.
Results Summary
| Method | Domain | N | EM | Δ |
|---|---|---|---|---|
| Baseline | Radiology | 950 | 46.4% | — |
| ★ RAG | Radiology | 950 | 60.6% | +14.2% |
| VQA-RAD (OOD) | Radiology | 451 | 67.8% | +24.6% |
| SLAKE (ID) | Radiology | 499 | 54.1% | +4.8% |
| RAG on PathVQA | Pathology | 500 | 36.4% | -3.4% |
Structure
medshift/ Core library (VLM wrapper, RAG engine, CCD)
experiments/ Experiment scripts
results/ Source center and memory bank
reports/ PDF reports
Quick Start
pip install -r requirements.txt
python experiments/run_radiology.py
python experiments/run_pathvqa.py
Citation
If you use this code, please cite our work.
Similar repos
Languages
Python
100.0%
README
MedShift
Training-free domain generalization for medical VLM hallucination mitigation via retrieval-augmented decoding.
Results Summary
| Method | Domain | N | EM | Δ |
|---|---|---|---|---|
| Baseline | Radiology | 950 | 46.4% | — |
| ★ RAG | Radiology | 950 | 60.6% | +14.2% |
| VQA-RAD (OOD) | Radiology | 451 | 67.8% | +24.6% |
| SLAKE (ID) | Radiology | 499 | 54.1% | +4.8% |
| RAG on PathVQA | Pathology | 500 | 36.4% | -3.4% |
Structure
medshift/ Core library (VLM wrapper, RAG engine, CCD)
experiments/ Experiment scripts
results/ Source center and memory bank
reports/ PDF reports
Quick Start
pip install -r requirements.txt
python experiments/run_radiology.py
python experiments/run_pathvqa.py
Citation
If you use this code, please cite our work.
Similar repos
Languages
Python
100.0%