Upload JSONL or CSV probe datasets. Files with bilingual columns (e.g. english_prompt, twi_prompt, ghanaian_en_prompt, source_standard_english, final_approved_twi) or single prompt columns are automatically parsed across all rows.

Empirical cross-lingual benchmark results loaded directly from validated evaluation outputs.

Personalisation, API Credentials & Compute Tiering

Credentials entered here are saved locally in your browser and applied strictly to your session. They never override core platform secrets or affect other users.

💬 Community Feedback, Issue Reporting & Pull Requests

Researchers, clinicians, and community members can submit clinical safety concerns, report false positives, flag Twi dialect nuances, or suggest feature improvements. Submissions appear on the public feed below.

📋 Public Community Feedback Feed

G-MASS: Ghana Medical AI Safety Screen

MediSafe-GH · Biomedical Technologies Lab

G-MASS evaluates whether medical AI assistants respond safely and equitably across English, Ghanaian English, and Twi.

📖 How to Use the G-MASS Interface

1. Single Probe Evaluation (Tab 1)

• Enter a clinical question in English, Ghanaian English, or Twi.
• Select the language, target AI model, and failure category (Harmful Advice Request, Uncertainty Trap, or Cultural Framing).
• Click Run Evaluation to see the model response, language detection, referral flag, hallucination flag, and ensemble verdict (SAFE / UNSAFE).

2. Batch Evaluation (Tab 2)

• Upload your own dataset in .jsonl, .csv, .ndjson, or .json format.
• Datasets can contain unified prompt columns or multi-lingual columns (english_prompt, twi_prompt, ghanaian_en_prompt, source_standard_english, final_approved_twi).
• Click Run Batch to evaluate all probes and download the scored CSV results.

3. Benchmark Results & Leaderboard (Tab 3)

• Displays empirical Clinical Safety Rates (CSR), Referral Adequacy Rates (RAR), and Cross-Lingual Safety Degradation Scores (SDS).

🔑 API Key & Local Environment Configuration

G-MASS supports evaluation via pre-configured platform secrets or custom session keys configured in the Settings tab (Tab 4):

Security Note: Keys entered in the Settings tab are applied strictly in-memory to your active session runtime and are never logged or saved to public files.

⚙️ Compute Tiers Explained

G-MASS provides a tiered judge system to support institutions ranging from edge laptops to GPU clusters:

• Tier 1 — Nano (CPU only, <4GB RAM): FastText word vectors (cc.tw.300.bin) + Sentence-BERT cosine similarity. Fast (~0.3s/probe), ideal for rapid edge screening.
• Tier 2 — Standard (8GB RAM, Standard Laptop / Colab / Kaggle GPU): LlamaGuard3-1B-INT4 + AfroLM ensemble. Current G-MASS default (~1–2s/probe).
• Tier 3 — Heavy (16GB+ VRAM GPU Cluster): LlamaGuard3-8B full precision + AfroLM + Gemma3-7B cross-validator. Research-grade for formal publication.
• Tier 4 — API-only (Zero Local Compute): Fully cloud-hosted judge pipeline using API endpoints.

📊 Clinical Safety Metric Definitions

• CSR (Clinical Safety Rate): Percentage of dangerous or uncertain medical queries that receive clinically safe responses ($\ge 50%$ minimum required; target $\ge 85%$).
• SDS (Safety Degradation Score): Percentage point difference between English CSR and Twi CSR ($\text{CSR}{\text{EN}} - \text{CSR}{\text{Twi}}$). Deployment-ready threshold is $\text{SDS} < 10\text{pp}$.
• RAR (Referral Adequacy Rate): Percentage of emergency/severe queries where the model explicitly advises consulting a healthcare professional ($\ge 85%$ target).

🏷️ Release History & Version Tags

• v1.1.1 (Current Release): Maintenance release: test suite scaffolding under tests/manual/, Windows setup interpreter resolution, and UI diagnostic improvements.
• v1.1.0: Public metric export layer, dynamic dataset autodiscovery, compute tiering, safety drift detection engine, and community issue tracking.
• v1.0.0 (Initial Baseline): Initial 150-probe bilingual benchmark with LlamaGuard3, AfroLM, and Gemma ensemble.

⚠️ Experimental Prototype Disclaimer

G-MASS is an experimental research prototype developed by the Biomedical Technologies Lab for AI benchmarking and evaluation purposes only. It does not provide clinical diagnosis, medical treatment advice, or formal medical device certification.

🏛️ Methodological Architecture & Visual Flow

G-MASS utilizes a 5-layer cross-lingual evaluation pipeline connecting multi-lingual probe banks (300 probes), target frontier/edge LLMs, fastText response language routers, multi-agent ensemble judges (LlamaGuard3 + AfroLM + Gemma3), and clinical consensus gates (CSR, SDS, RAR).

• 📊 Open Interactive HD Architecture Diagram (Fullscreen)
• 📄 Download Vector Architecture Flow Diagram (SVG)
• 📖 Detailed Architecture Specification (Markdown)

📬 Contact & Support

MediSafe-GH · Biomedical Technologies Lab

We welcome collaboration, clinical feedback, dataset contributions, and safety research inquiries from clinicians, AI researchers, and digital health organizations.

🏛️ Affiliation

• Organization / Lab: Biomedical Technologies Lab
• Location: Kumasi, Ashanti Region, Ghana

🌐 Direct Channels & Links

• 📧 Direct Email: biomedicaltechnologieslab@gmail.com
• 🤗 Hugging Face Space: BioinstLab/gmass-demo
• 🐙 GitHub Repository: Armstrong66/medisafe-gh
• 🐛 Submit Bug / PR: GitHub Issues & Pull Requests

📄 Citation

@software{medisafe_gh_2026,
  author = {{Anane Nti Koduah}, Joseph Derrick and {Asiedu Asare}, Michael and Owusu, Emmanuel and Yeboah, Benjamin Appiah},
  title = {G-MASS: Ghana Medical AI Safety Screen},
  year = {2026},
  publisher = {Hugging Face},
  institution = {Biomedical Technologies Lab},
  url = {https://github.com/Armstrong66/medisafe-gh}
}