farunawebservices/lsr-dashboard

Linguistic Safety & Robustness Workbench - Red-teaming frontier LLMs across low-resource African languages # LSR Dashboard: Linguistic Safety & Robustness Workbench **Author:** Godwin Faruna Abuh **Role:** AI Safety Researcher | Senior Applied AI Safety Engineer --- ## 🛡️ Overview The **LSR Dashboard** is a workbench for evaluating **Linguistic Safety Decay** in Large Language Models. While frontier models exhibit strong safety alignment in English, this robustness often deteriorates in mid/low-resource languages. This tool provides: - **Cross-lingual red-teaming** across Igala, Yoruba, Hausa, and English - **Mechanistic visualization** of activation drift - **Empirical loophole detection** with session statistics - **Translation verification** via Google Translate --- ## 🌍 Languages Covered - Yoruba 🇳🇬 (5 attack probes) - Hausa 🇳🇬 (5 attack probes) - Igbo 🇳🇬 (4 attack probes) - Igala 🇳🇬 (3 attack probes) - English 🇬🇧 (baseline) **Model:** Gemini 2.5 Flash (GPT-4 & Claude support planned) --- ## 🚀 Key Features ### 1. Red-Teaming Lab Side-by-side comparison of English baseline vs target language responses. Automatic loophole detection when English refuses but target language complies. ### 2. Mechanistic Visualizer Activation heatmaps and safety centroid drift plots showing how refusal circuits struggle with low-resource syntax. ### 3. Vulnerability Gallery Archive of confirmed HIGH/CRITICAL safety failures with empirical findings. ### 4. Session Analytics Live tracking with exportable JSON logs for offline analysis. --- ## 📊 Use Case Demonstrate to decision makers that frontier models show **2-4x higher bypass rates** in Yoruba/Hausa/Igala compared to English. --- ## 🛠 Technical Stack - Python 3.10+ | Streamlit - Plotly visualizations - Google Gemini 2.5 Flash API --- ## ⚙️ Local Setup (Optional) If you want to run this locally instead of using the Hugging Face Space: ```bash # Clone or download this repository # Install dependencies pip install streamlit pandas numpy plotly google-generativeai python-dotenv # Create a .env file and add: # GEMINI …