josephgitau/The-African-Trust-Safety-LLM-Challenge

# African Trust & Safety LLM Challenge — Red-Team Tool A web-based tool for the Zindi African Trust & Safety LLM Challenge. Load African language LLMs locally, craft adversarial prompts, validate safety breaks, and export structured submissions. ## Features - **Local Model Loading** — Load African language LLMs (Swahili, Hausa, Yoruba, Amharic) directly on your GPU via HuggingFace Transformers - **Streaming Generation** — Real-time token streaming with SSE - **Refusal Detection** — Automatic detection of model refusals using multilingual keyword matching - **Validation Runs** — Run prompts multiple times (up to 10×) to confirm break reproducibility - **Break Database** — SQLite-backed storage for confirmed safety breaks with full CRUD - **Taxonomy System** — Attack types, risk categories, and subcategories from the challenge spec - **AI-Powered Features** (optional, requires OpenAI API key): - Prompt optimization - Auto-categorization of attacks - Contextual note generation - Translation between languages - **Export** — Generate Zindi-format markdown submissions from saved breaks ## Requirements - **Python 3.10+** - **NVIDIA GPU** with CUDA support (models run on GPU) - ~4-8 GB VRAM depending on the model ## Setup 1. **Clone the repo** ```bash git clone github.com cd The-African-Trust-Safety-LLM-Challenge ``` 2. **Install dependencies** ```bash pip install -r requirements.txt ``` 3. **Run the app** ```bash python run_webapp.py ``` Opens automatically at 127.0.0.1 ## Usage 1. **Load a model** — Select a model from the top bar dropdown and click "Load". The model downloads from HuggingFace on first use. 2. **Write a prompt** — Type your adversarial prompt in the Prompt Lab (in the target language). 3. **Run** — Click "Run" for streaming generation, or "Validate 3×" to confirm reproducibility. 4. **Save breaks** — When you find a safety failure, click "Save Break" to store it with m …