metadata
license: apache-2.0
configs:
- config_name: v20251015-bytes
data_files:
- split: GMD_complexQA
path: GMD_complexQA/publish-20251015.bytes.parquet
- split: GMD_simpleQA
path: GMD_simpleQA/publish-20251015.bytes.parquet
- split: LTR_abnormalityQA
path: LTR_abnormalityQA/publish-20251015.bytes.parquet
- split: LTR_fullparsing
path: LTR_fullparsing/publish-20251015.bytes.parquet
- split: LTR_simpleQA
path: LTR_simpleQA/publish-20251015.bytes.parquet
MedDocBench
A compact benchmark of text‑rich medical document understanding covering routine, patient-uploaded artifacts from online consultations, spanning laboratory test reports (LTR) and general medical documents (GMD).
Configuration and splits
- Configuration:
v20251015-bytes - Available splits:
- LTR_fullparsing: 100
- LTR_simpleQA: 200
- LTR_abnormalityQA (complex QA): 100
- GMD_simpleQA: 100
- GMD_complexQA: 100
- Total: 600 QA pairs
Data formats
publish-20251015.bytes.parquet: images stored as base64-encoded bytes with relative paths.publish-20251015.parquet: images referenced via relative paths (no embedded bytes).tsvfiles: images referenced via relative paths, for direct evaluation with EvalScope/VLMEvalKit.
Statistics
Angle distribution (all tasks): 0°/90°/180°/270° = 25%/25%/25%/25%
LTR image distribution
| Status | Capture method | Count |
|---|---|---|
| Normal | Mobile | 4 |
| Normal | Paper | 16 |
| Abnormal | Mobile | 16 |
| Abnormal | Paper | 64 |
GMD document type distribution
- GMD_simpleQA
| Doc type | Count |
|---|---|
| Laboratory Test Reports | 30 |
| Medication Packages | 19 |
| Imaging Reports | 11 |
| Outpatient Encounter Notes | 9 |
| Other Diagnostic Reports | 9 |
| Inpatient Records | 7 |
| Prescriptions (Western Medicine) | 6 |
| Other Clinical Documents | 4 |
| Prescriptions (TCM) | 4 |
| Health Records | 1 |
- GMD_complexQA
| Doc type | Count |
|---|---|
| Medication Packages | 24 |
| Imaging Reports | 15 |
| Other Diagnostic Reports | 13 |
| Laboratory Test Reports | 13 |
| Prescriptions (Western Medicine) | 10 |
| Outpatient Encounter Notes | 7 |
| Inpatient Records | 5 |
| Other Clinical Documents | 4 |
| Prescriptions (TCM) | 4 |
| Other text-rich images | 3 |
| Health Records | 2 |
Load with datasets
from datasets import load_dataset
# Load the HuggingFace dataset and pick the configuration
ds = load_dataset("<ORG_OR_USER>/MedDocBench", "v20251015-bytes")
# Access individual splits by name
gmd_simple = ds["GMD_simpleQA"]
ltr_full = ds["LTR_fullparsing"]
Evaluation
Based on EvalScope with a VLMEvalKit backend. See the EvalScope documentation: evalscope.readthedocs.io.
# 1) Create env (Python 3.10 recommended)
conda create -n evalscope python=3.10 -y
conda activate evalscope
# 2) Install EvalScope with VLMEvalKit extras
pip install "evalscope[vlmeval]"
# 3) Point EvalScope to your TSVs used by benchmark_vlmevalkit
# Create or edit the .env file under the installed package, e.g.:
# $CONDA_PREFIX/lib/python3.10/site-packages/.env
# and add:
# LMUData=/path/to/tsv_files
# 4) Run evaluation from the benchmark folder
cd MedDocBench/evaluation
python eval.py --config eval_config.yaml
Citation
If you use this benchmark, please cite:
Citrus‑V: Advancing Medical Foundation Models with Unified Medical Image Grounding for Clinical Reasoning
arXiv:2509.19090
License
license: apache-2.0