Dataset Card for MamaBench

Dataset Summary

MamaBench is a counterfactual clinical benchmark for evaluating the robustness of large language models on maternal and child health diagnostic reasoning. It consists of 217 counterfactual case pairs, each pairing an original clinical vignette with a systematically perturbed counterfactual variant, designed to test whether a model's diagnostic reasoning is sensitive to clinically meaningful changes rather than relying on surface-level pattern matching.

MamaBench is released as a single, unsplit evaluation set. It is intended exclusively for zero-shot evaluation — see Data Splits below.

Supported Tasks

• Diagnostic robustness evaluation: measuring whether a model's diagnosis changes appropriately (or fails to change) in response to clinically significant counterfactual perturbations.
• Benchmarking mitigation architectures (e.g., retrieval-augmented or evidence-anchored pipelines) against standalone model baselines.

Languages

English.

Dataset Structure

Data Instances

Each pair consists of two rows sharing the same pair_id — one case presenting the original clinical scenario, and one presenting a counterfactually perturbed variant. The two narratives share overlapping surface features but differ in clinically decisive details, leading to different ground-truth diagnoses.

Data Fields

• case_id: string — unique identifier for an individual case
• pair_id: string — shared identifier linking a case to its counterfactual partner; two case_ids share one pair_id
• narrative: string — first-person clinical presentation narrative
• ground_truth: string — the correct diagnosis for that specific narrative

Note: MamaBench does not store a static PP/PF/FP/FF label as a data field — that taxonomy is a derived evaluation output (whether a model gets the original and counterfactual case in a pair right or wrong), computed at scoring time, not part of the released schema.

Data Splits

MamaBench has no train/validation/test split. All 217 pairs are released as a single evaluation set.

This is intentional: MamaBench is designed as a held-out, zero-shot benchmark. Using any portion of this dataset for fine-tuning, few-shot prompting, or in-context training will invalidate its use as an independent evaluation of model robustness. If you are developing or tuning a model, do not train on MamaBench.

Dataset Creation

Curation Rationale

Existing clinical LLM benchmarks largely evaluate static diagnostic accuracy and do not test whether a model's reasoning is robust to clinically meaningful changes in presentation — a gap that matters directly for safe deployment of LLM-based clinical decision support across healthcare settings generally, not any single region or resource context.

Source Data

Case scenarios were authored by a team of clinician co-investigators, drawing on their clinical expertise. No real patient records, PHI, or identifiable patient information were used at any stage.

Annotations

Each case's ground_truth diagnosis was assigned by the clinician co-investigators as part of authoring the narrative. The PP/PF/FP/FF taxonomy referenced elsewhere in this card is not an annotation on the data — it is computed at evaluation time from a model's outputs on a pair, not part of the dataset creation process.

Personal and Sensitive Information

This dataset contains no personal, identifiable, or patient-derived information. All clinical scenarios are synthetic, authored by clinician co-investigators rather than drawn from real patient records.

Considerations for Using the Data

Intended Use

MamaBench is intended to support the evaluation and development of safer LLM-based clinical decision support systems for maternal and child health, across healthcare settings broadly.

Limitations

• The benchmark is intentionally small (217 pairs) to allow high-quality expert curation over scale; it is not intended as a training corpus.
• Clinical scenarios are authored to reflect general clinical practice and guidelines and may not capture every regional or institutional variation in care.
• As with any fully public static benchmark, contamination risk increases over time as the data is scraped into future pretraining corpora (see note below).

Additional Information

Licensing Information

This dataset is released under the CC BY 4.0 license.

Citation Information

      title={MamaBench: Benchmarking LLM Robustness in Maternal and Child Health Diagnosis through Counterfactual Clinical Perturbation}, 
      author={Thanni Adewuyi and Anuoluwa Sotome and Samuel Okoko and Angel Ezendu and Oluwafunke Akinbuwa and Oluwaseun Odunsi and Oluwasegun Oguntuase and Oluwadarasimi Oguntuase and Ifeoma Nwabueze and Abiodun Adereni},
      year={2026},
      eprint={2607.14385},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2607.14385}, 
}

Ethics Statement

This study did not involve human subjects research. Case scenarios were authored by co-investigators drawing on clinical expertise; no patient data, identifiable information, or external human participants were involved.

Contributions

Thanks to the clinician co-investigators who authored the benchmark scenarios, and to the HelpMum Africa AI/ML team.