---
license: apache-2.0
---


# MedSR-Bench

### A Full-PRISMA-Workflow-Supporting Benchmark for Medical Systematic Review Automation

MedSR-Bench is released with **MedSR-Copilot**, a PRISMA-aligned agentic framework for automated medical systematic reviews. It evaluates systems from a structured review protocol to final evidence-synthesis conclusions.

---

## Dataset Summary

MedSR-Bench contains **100 systematic reviews across 24 medical domains**. The benchmark includes **1,297 included RCT records**, **2,144 predefined analysis groups**, **7,570 samples for data extraction** and **1,908 samples for end-to-end conclusion evaluation**.

This release contains two complementary components:

- `MedSR-Bench-json/`: 100 structured SR records with labels for literature retrieval, screening, data extraction, Risk-of-Bias (RoB) assessment, and evidence synthesis.
- `SR-RoB-7B-dataset/`: Parquet files for training and evaluating **SR-RoB-7B** on Cochrane Risk-of-Bias tool.

## Highlights

- 📚 **SR-grounded** — derived from systematic reviews with traceable protocols, included studies, extraction records, and RoB judgments.
- 🔄 **End-to-end** — supports literature retrieval, screening, data extraction, RoB assessment, and statistical evidence synthesis in one benchmark.
- 📊 **Scale & breadth** — 100 systematic reviews spanning 24 medical domains, with 2,144 predefined analysis groups.
- 🎯 **Protocol-aligned evaluation** — every review starts from a PICOS-based protocol and detailed eligibility criteria.
- 🩺 **Dedicated RoB resource** — includes training and test datasets for SR-RoB-7B, covering five RoB 1.0 domains.

## Dataset Structure

```text
MedSR-Bench/
├── MedSR-Bench-json/
│   ├── 1.json
│   ├── ...
│   └── 100.json
└── SR-RoB-7B-dataset/
    ├── train.parquet
    └── test.parquet
```

## Usage

Each JSON file represents one systematic review. `PICO` stores the PICOS protocol and eligibility criteria; `study group` stores predefined analysis groups and the final `gt conclusion`; `trials` provides the gold-standard included studies for retrieval and screening. Study named after individual studies contain the corresponding data-extraction and RoB labels.

The evidence-synthesis label is one of `favor Intervention`, `favor comparison`, or `not significant`.

`SR-RoB-7B-dataset/` provides Parquet data for SR-RoB-7B training and evaluation. Each record is an included study with a full-text-based RoB task and labels for five RoB domains.

```python
from huggingface_hub import snapshot_download

path = snapshot_download(
    repo_id="YOUR_NAMESPACE/MedSR-Bench",
    repo_type="dataset",
)
```

## Source

MedSR-Bench was constructed from open-access Cochrane systematic reviews archived in PMC. 

## Citation

<!-- To be added. -->

## License

Released under the apache-2.0 License.