shizaamir1615/human-vs-machine-text-detector

Group 23- Detecting machine generated content in low resource African languages Human vs Machine Text Detector For South African Civic/Government Text in Low-Resource Languages This Streamlit application leverages an ensemble of two fine-tuned transformer models—AfriBERTa and XLM-RoBERTa—to classify whether a given sentence in a civic or governmental context is human-written or machine-generated. The tool is designed to support four low-resource South African languages: Xhosa (xho) Tsonga (tso) Tshivenda (ven) Northern Sotho (nso) Features Ensemble Predictions: Combines outputs from AfriBERTa and XLM-RoBERTa for robust classification. LIME Explanations: Provides transparent, word-level insights into model predictions. Multi-Language Support: Handles four low-resource South African languages. Custom UI: Features a sleek, dark-themed interface for enhanced user experience. Offline Capability: Models are loaded from local directories, ensuring no internet dependency. Robust Error Handling: Gracefully manages issues during model loading and inference. Folder Structure project-root/ │ ├── app.py # Streamlit demo interface ├── model_utils.py # Shared ensemble prediction functions ├── evaluate_model.py # Evaluation pipeline for metrics and robustness ├── requirements.txt # Project dependencies ├── afriberta_dir/ # Fine-tuned AfriBERTa model directory ├── xlmr_dir/ # Fine-tuned XLM-RoBERTa model directory └── final_dataset2.csv # Cleaned dataset for evaluation Each model directory (afriberta_dir/ and xlmr_dir/) must contain: config.json pytorch_model.bin or model.safetensors tokenizer_config.json, tokenizer.json Additional tokenizer files: vocab.txt, special_tokens_map.json, etc. Installation Create and activate a virtual environment: python3 -m venv venv source venv/bin/activate # On Windows: venv\Scripts\activate Install dependencies: pip install -r requirements.txt Running the App Start the Streamlit app: streamlit run app.py Open your browser and navigate to …