Datasets:

NYU-OLAB
/

CABA

---
license: cc-by-4.0
pretty_name: Clinical AI Benchmark Atlas (CABA)
language:
  - en
size_categories:
  - 10K<n<100K
tags:
  - medical
  - clinical-ai
  - benchmarks
  - llm-evaluation
configs:
  - config_name: registry
    data_files: registry.csv
  - config_name: per_benchmark
    data_files: per_benchmark.csv
  - config_name: labels
    data_files: labels.csv
---
# Clinical AI Benchmark Atlas (CABA)

CABA is a registry of the public benchmarks used to evaluate large language models in health care: the 58
benchmarks named by 13 evaluation suites and by frontier model reports, checked against five criteria on their
downloaded items. Up to 200 items from each carry [RCQ-Map](https://github.com/nyuolab/RCQ-Map) labels, and
every benchmark is compared with the task mix of real clinical queries that clinicians sent to an institutional
AI assistant. Version 1.0 is the version described in the paper.

**Benchmark items are not included.** Each benchmark keeps its own license, and some ask that their items not be
posted online. CABA releases item ids and labels; the scripts in the [GitHub repository](https://github.com/nyuolab/CABA)
rebuild the items from each benchmark's own release.

## Files

| Config | Rows | Contents |
|---|---|---|
| `registry` | 58 | One row per benchmark: evaluation suites, frontier model reports, items sampled and labeled, split, version, license, source, rendering rule |
| `per_benchmark` | 58 | Task-category and kind-of-work shares per benchmark, and overlap with the real clinical task mix |
| `labels` | 10,659 | One row per benchmark item: `item_id`, `benchmark` and the 24 RCQ-Map fields. No item text |

`rcq_aggregate.json` holds the aggregate shares of the 127,625 real clinical queries used for comparison.

```python
from datasets import load_dataset
labels = load_dataset("NYU-OLAB/CABA", "labels", split="train")
```

## How the labels were made

Up to 200 items per benchmark were sampled, rendered as a model receives them, and labeled with RCQ-Map by the
paper's annotator (GPT-5.6 sol, reasoning effort none, the paper's guidelines, strict JSON schema). The 40 sampled
Med-HALT items that failed the schema's consistency rules are listed in the GitHub repository. The protocol, the
registry build stages and the comparison script are there as well, and `scripts/compare.py` reproduces the paper's
medians from these labels.

## License and citation

Registry, labels, summaries and reference shares: CC BY 4.0. The benchmarks themselves keep their own licenses,
recorded in `registry`. If you use CABA, please cite the paper (Vishwanath et al., 2026; see the GitHub
repository's `CITATION.cff`).

	---
	license: cc-by-4.0
	pretty_name: Clinical AI Benchmark Atlas (CABA)
	language:
	- en
	size_categories:
	- 10K<n<100K
	tags:
	- medical
	- clinical-ai
	- benchmarks
	- llm-evaluation
	configs:
	- config_name: registry
	data_files: registry.csv
	- config_name: per_benchmark
	data_files: per_benchmark.csv
	- config_name: labels
	data_files: labels.csv
	---

	# Clinical AI Benchmark Atlas (CABA)

	CABA is a registry of the public benchmarks used to evaluate large language models in health care: the 58
	benchmarks named by 13 evaluation suites and by frontier model reports, checked against five criteria on their
	downloaded items. Up to 200 items from each carry [RCQ-Map](https://github.com/nyuolab/RCQ-Map) labels, and
	every benchmark is compared with the task mix of real clinical queries that clinicians sent to an institutional
	AI assistant. Version 1.0 is the version described in the paper.

	Benchmark items are not included. Each benchmark keeps its own license, and some ask that their items not be
	posted online. CABA releases item ids and labels; the scripts in the [GitHub repository](https://github.com/nyuolab/CABA)
	rebuild the items from each benchmark's own release.

	## Files

	\| Config \| Rows \| Contents \|
	\|---\|---\|---\|
	\| `registry` \| 58 \| One row per benchmark: evaluation suites, frontier model reports, items sampled and labeled, split, version, license, source, rendering rule \|
	\| `per_benchmark` \| 58 \| Task-category and kind-of-work shares per benchmark, and overlap with the real clinical task mix \|
	\| `labels` \| 10,659 \| One row per benchmark item: `item_id`, `benchmark` and the 24 RCQ-Map fields. No item text \|

	`rcq_aggregate.json` holds the aggregate shares of the 127,625 real clinical queries used for comparison.

	```python
	from datasets import load_dataset
	labels = load_dataset("NYU-OLAB/CABA", "labels", split="train")
	```

	## How the labels were made

	Up to 200 items per benchmark were sampled, rendered as a model receives them, and labeled with RCQ-Map by the
	paper's annotator (GPT-5.6 sol, reasoning effort none, the paper's guidelines, strict JSON schema). The 40 sampled
	Med-HALT items that failed the schema's consistency rules are listed in the GitHub repository. The protocol, the
	registry build stages and the comparison script are there as well, and `scripts/compare.py` reproduces the paper's
	medians from these labels.

	## License and citation

	Registry, labels, summaries and reference shares: CC BY 4.0. The benchmarks themselves keep their own licenses,
	recorded in `registry`. If you use CABA, please cite the paper (Vishwanath et al., 2026; see the GitHub
	repository's `CITATION.cff`).