---
pretty_name: BenchBase TRIAGE
license: other
task_categories:
  - question-answering
language:
  - en
tags:
  - medical
  - clinical
  - benchmarking
  - multiple-choice
source_datasets:
  - NLie2/TRIAGE
---

# BenchBase TRIAGE

TRIAGE normalized into the BenchBase unified schema. Every BenchBase dataset shares the same
columns, so one loader and one evaluation loop cover all of them.

- **Source:** [NLie2/TRIAGE](https://huggingface.co/datasets/NLie2/TRIAGE)
- **Code and converter:** [Layered-Labs/benchbase-med](https://github.com/Layered-Labs/benchbase-med)
- **Items:** 86 (train: 86)
- **Rejected source rows:** 1 (see `rejected.json`)

## Schema

| Field | Type | Description |
|---|---|---|
| `dataset_key` | str | Benchmark identifier (`triage`) |
| `hash` | str | SHA256(question + answer text), stable across option order |
| `split` | str | Source split |
| `question_type` | str | `mcq` |
| `question` | str | Question stem |
| `context` | str or null | Passage the question refers to, if any |
| `options` | list[dict] | `[{"original_key": "A", "text": "..."}]` |
| `answer` | dict | The matching option: `{"original_key": "D", "text": "..."}` |
| `metadata` | dict | Source fields that do not fit the schema |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("Layered-Labs/benchbase-triage")
print(ds["train"][0])
```