Stella-Achar-Oiro/gender-sensitization-engine

JuaKazi Gender Sensitization Engine — bias detection & correction for African languages (EN/SW/FR/KI/ZU/HA) --- title: JuaKazi Gender Sensitization Engine emoji: ⚖️ colorFrom: red colorTo: green sdk: docker app_port: 7860 pinned: false --- # JuaKazi Gender Sensitization Engine A multilingual gender bias detection and correction system for African language text, built for the AI BRIDGE programme. ## What it does Detects and rewrites gender-biased language in **Swahili, English, French, Gikuyu, Hausa, and Zulu**. Supports single-sentence and batch API input. ## Current metrics (May 2026) | Language | F1 | Precision | Recall | Samples | |---|---|---|---|---| | English | 1.000 | 1.000 | 1.000 | 66 | | Swahili | 0.851 | 0.822 | 0.881 | 67,290 | | French | 0.970 | 1.000 | 0.941 | 165 | | Gikuyu | 0.667 | 0.967 | 0.510 | 11,622 | | Hausa | 0.043 | 1.000 | 0.022 | 10,054 | | Zulu | 0.732 | 1.000 | 0.577 | 2,000 | Hausa: precision-first initial lexicon (36 rules); recall requires ML classifier (planned). Zulu: morphological suffix rules with context gating for zero false positives. ## Live demo huggingface.co ## ML models | Model | Base | Val F1 | Val Precision | Val Recall | Notes | |---|---|---|---|---|---| | sw-bias-classifier-v1 | afro-xlmr-base | 0.854 | 0.938 | 0.784 | Full fine-tune, 51K rows | | sw-bias-classifier-v2 | afro-xlmr-base | 0.953 | 0.940 | 0.960 | Overfit on val — invalid | | sw-bias-classifier-v3 | afro-xlmr-base | 0.871 | 0.810 | 0.942 | Current deployed model | SW ML classifier is Stage 2 fallback only — runs when rules find nothing. ## Architecture - **Detection**: Deterministic lexicon rules (primary) + AfroXLM-R fine-tuned ML fallback (Swahili only) - **Correction**: Word-level substitution with semantic preservation check (threshold 0.70) - **Languages**: Separate lexicons per language — no cross-lingual transfer - **Context gating**: 11 suppression conditions (biographical, quote, statistical, counter-stereotype, Zulu neutral profession, etc.) - **Integration**: `caller` field allows ups …