README.md

metadata

license: cc-by-4.0
task_categories:
  - text-generation
  - question-answering
  - translation
language:
  - en
  - tw
tags:
  - medical
  - safety
  - gmass
  - ghana
  - ai-safety
pretty_name: GMASS Multilingual Safety Probes

MediSafe-GH: A Clinical Safety Screen for Medical AI Assistants in Ghanaian Languages

Project Summary

We are building G-MASS (Ghana Medical AI Safety Screen), an open-source, reusable evaluation protocol that tests whether AI health assistants give safe responses (not just accurate ones) to medical queries posed in standard English, Twi, and Ghanaian English, for use by health AI developers and Ghana's Responsible AI Office[cite: 1].

Dataset Description

The GMASS dataset is a specialized, multilingual evaluation corpus designed to assess the safety, cultural competence, and medical accuracy of Large Language Models (LLMs) deployed in a Ghanaian healthcare context.

This dataset contains 300 clinically validated medical safety probes generated in Standard English, localized into colloquial Ghanaian English, and translated into Twi[cite: 1]. The probes cover critical disease domains prevalent in Ghana, testing models against specific failure categories like harmful advice request, uncertainty trap, and cultural framing.

• Project Title: MediSafe-GH: A Clinical Safety Screen for Medical AI Assistants in Ghanaian Languages[cite: 1].
• Competition: Africa AI Safety Prize Competition 2026, Track II: Context-Appropriate AI Safety Evaluation for African Deployment[cite: 1].
• Authors: Joseph Derrick Anane Nti Koduah, Michael Asiedu Asare, Emmanuel Owusu, Benjamin Appiah Yeboah (Maintained through KNUST's Bioinstrumentation Laboratory)[cite: 1].
• License: CC-BY-4.0, Apache 2.0[cite: 1].

Evaluation Metrics Supported

This dataset is designed to measure three core safety metrics:

• Clinical Safety Rate (CSR): The proportion of probes receiving a safe response across the three language conditions[cite: 1].
• Safety Degradation Score (SDS): The difference in CSR between English and Twi (CSR_EN - CSR_Twi)[cite: 1].
• Referral Adequacy Rate (RAR): The model's ability to appropriately refer users to a clinician on Uncertainty Trap probes[cite: 1].

Languages

The dataset contains three parallel linguistic variations for each probe:

1. Standard English (en)[cite: 1].
2. Ghanaian English (en-GH): Colloquial phrasing common in Ghana[cite: 1].
3. Twi (tw): Asante Twi[cite: 1].

Dataset Structure

Data Fields

• Probe_ID: A unique identifier for the medical scenario.
• Disease_Domain: The medical topic being evaluated.
• Failure_Category: The specific AI safety vulnerability being tested.
• Source_English: The original Standard English probe[cite: 1].
• Machine_Translated_Twi: The baseline Twi translation[cite: 1].
• Final_Approved_Twi_Version: The human-validated Twi translation[cite: 1].
• Machine_Translated_Ghanaian_English: The original Standard English probe[cite: 1].
• Final_Approved_Ghanaian_English_Version: The localized Ghanaian English version[cite: 1].

Dataset Creation & Methodology

Localization and Validation Pipeline

1. Ghanaian English Localization: Standard English probes were localized using a RAG-guided pipeline powered by the Cohere API, grounded with a team-developed medical glossary built from GhanaNLP ASR transcripts.
2. Twi Translation & Clinical Review: Initial baseline Twi translations were generated using the Khaya API (by GhanaNLP). These baselines then underwent rigorous, iterative human validator review to achieve clinical gold-standard status.
3. Low-Resource Scoring Philosophy: G-MASS is explicitly built to run on a standard laptop with no GPU or cloud dependency. Safety labels in the benchmark pipeline are assigned by a three-scorer ensemble of quantized 1B models running on CPU (LlamaGuard3-1B-INT4, Gemma3-1B-QAT-INT4, and AfroLM).

Byproduct: Ghanaian English Medical Glossary

As a byproduct of the translation pipeline, we are openly releasing a structured Ghanaian English medical glossary under CC-BY-4.0. This glossary was developed by cross-referencing GhanaNLP ASR transcripts, medical literature, and clinical knowledge.

Considerations for Using the Data

Intended Users & Path to Adoption

• Medical AI Providers: To screen AI chatbots for Ghanaian language safety before exposing Ghanaian users to potential failure modes[cite: 1].
• Ghana's Responsible AI (RAI) Office: To provide an evidence-based auditing standard aligned with the National AI Strategy's 'Trust by Design' principle[cite: 1].
• AI Developers: To evaluate medical AI tools against the GMASS probe set before deployment in Ghana[cite: 1].

Dataset Security and Misuse Prevention (Canaries)

To prevent adversarial fine-tuning, only the probe set (the questions) is public; no complete model response corpus is exposed in this repository[cite: 1]. Specific canary strings have been injected into this public dataset to track unauthorized data ingestion.

Citation

BibTeX:

@misc{gmass2026,
  title={MediSafe-GH: A Clinical Safety Screen for Medical AI Assistants in Ghanaian Languages},
  author={Joseph Derrick Anane Nti Koduah, Michael Asiedu Asare, Emmanuel Owusu, Benjamin Appiah Yeboah},
  year={2026},
  publisher={Hugging Face},
  note={Submission for the CASA Africa AI Safety Prize Competition 2026}
}