|
|
--- |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- en |
|
|
tags: |
|
|
- medical |
|
|
- benchmark |
|
|
- evaluation |
|
|
- clinical-guidelines |
|
|
- diagnostic-reasoning |
|
|
pretty_name: AMEGA-LLM Benchmark |
|
|
task_categories: |
|
|
- question-answering |
|
|
configs: |
|
|
- config_name: cases |
|
|
data_files: "cases.csv" |
|
|
sep: ";" |
|
|
default: true |
|
|
- config_name: questions |
|
|
data_files: "questions.csv" |
|
|
sep: ";" |
|
|
- config_name: sections |
|
|
data_files: "sections.csv" |
|
|
sep: ";" |
|
|
- config_name: criteria |
|
|
data_files: "criteria.csv" |
|
|
sep: ";" |
|
|
|
|
|
extra_gated_heading: "Evaluation-only access" |
|
|
extra_gated_description: "Do NOT use this dataset for model training, pretraining, fine-tuning, or data augmentation." |
|
|
extra_gated_prompt: "By requesting access, you agree to all of the following:" |
|
|
extra_gated_button_content: "Agree and request access" |
|
|
extra_gated_fields: |
|
|
I will not use this dataset for training/pretraining/fine-tuning/data augmentation: checkbox |
|
|
I will not redistribute the Q/A content: checkbox |
|
|
Intended use (one line): text |
|
|
Affiliation (optional): text |
|
|
--- |
|
|
|
|
|
# AMEGA-LLM Benchmark |
|
|
|
|
|
20 guideline-based clinical cases across 13 specialties with open-ended questions and a detailed rubric (1,337 criteria) to evaluate LLM medical reasoning and **adherence to clinical guidelines**. |
|
|
|
|
|
> **Evaluation-only — do not use for training.** |
|
|
|
|
|
## Files / configs |
|
|
- `cases` – 20 case narratives + metadata |
|
|
- `questions` – all questions per case |
|
|
- `sections` – rubric sections (with point weights) |
|
|
- `criteria` – fine-grained checklist items |
|
|
|
|
|
## Quick start |
|
|
```python |
|
|
from datasets import load_dataset |
|
|
|
|
|
# Load the cases table |
|
|
cases = load_dataset("row56/amega-benchmark", "cases") |
|
|
print(cases["train"][0]) |
|
|
|
|
|
# Load other tables |
|
|
questions = load_dataset("row56/amega-benchmark", "questions") |
|
|
sections = load_dataset("row56/amega-benchmark", "sections") |
|
|
criteria = load_dataset("row56/amega-benchmark", "criteria") |
|
|
``` |
|
|
|
|
|
## Intended use & canary |
|
|
This dataset is intended for benchmarking/evaluation of LLM clinical reasoning and guideline adherence. **Do not** fine-tune or train models on this dataset. |
|
|
A unique canary marker is embedded in the repository to help detect misuse; please leave it intact and do not reproduce it in documentation or prompts. |
|
|
|
|
|
## Citation |
|
|
Fast et al., *npj Digital Medicine* (2024). |
|
|
|
|
|
```bibtex |
|
|
@article{fast2024amega, |
|
|
title={Autonomous Medical Evaluation for Guideline Adherence of LLMs}, |
|
|
journal={npj Digital Medicine}, |
|
|
year={2024} |
|
|
} |
|
|
``` |
|
|
|
|
|
## Version |
|
|
- v1.0 — initial release (Aug 10, 2025) |
|
|
|
|
|
## Links |
|
|
- Paper (open access): https://doi.org/10.1038/s41746-024-01356-6 |
|
|
- Source code / issues: https://github.com/DATEXIS/AMEGA-benchmark |
|
|
- Canary marker file: `CANARY.txt` |
