The premier medical safety benchmark for chatbots/LLMs in Ghana.
# MediSafe-GH - G-MASS
**Ghana Medical AI Safety Screen** -- an open-source evaluation pipeline for
testing whether medical AI assistants behave safely when queried in English,
Twi, and Ghanaian English.
Africa AI Safety Prize Competition 2026 - Track II submission.
## What The Pipeline Does
G-MASS evaluates model responses with clinical safety metrics, not just general
helpfulness or factuality.
- **CSR**: Clinical Safety Rate, the percentage of responses scored `SAFE`.
- **SDS**: Safety Degradation Score, computed as CSR_English minus CSR_Twi or CSR_GH-EN.
- **RAR**: Referral Adequacy Rate, the percentage of crisis-level probes where the model refers the user to a clinician.
The active entry point is:
```powershell
gmass ...
```
## Probe-Tested Models
The pipeline currently evaluates these 4 probe-tested models:
| Key | Model ID | Provider/backend |
|---|---|---|
| `gpt4o` | `gpt-4o` | OpenAI |
| `gemini` | `gemini-2.5-flash` by default, overrideable with `GEMINI_MODEL` | Google GenAI |
| `phi3` | `microsoft/Phi-3-mini-4k-instruct` | Local `transformers`, local OpenAI-compatible server, or HF router |
| `biomistral` | `BioMistral/BioMistral-7B-SLERP` | Local `transformers`, local OpenAI-compatible server, or HF router |
Scorers are not probe-tested models. Do not count these as evaluated models:
- `LlamaGuard3`
- `AfroLM`
- `Gemma`
- `meta-llama/Llama-Guard-3-1B`
- `google/gemma-3-1b-it`
`gemini` can be an evaluated model key. Separately, `SCORER_BACKEND=policy_api`
means the scorer pipeline uses a hosted policy-judge runtime to execute scorer
prompts. That runtime may currently call Gemini API under the hood, but the
scorer identities remain `LlamaGuard3`, `Gemma`, and `AfroLM`.
The `gemini` key is a stable pipeline key for a Gemini Flash model, not a claim
that the historical `gemini-1.5-flash` endpoint is still available. Current
defaults use `gemini-2.5-flash`.
For extension notes, see `docs/model-and-scorer-extensibility.md`.
## Environment Se …