gellebereed/Somali-BBC-News-Summarization

# Somali BBC News Summarization (FLAN-T5 + LoRA) This repository contains a complete end-to-end NLP system for **abstractive summarization of Somali BBC news articles**, developed as a final project for a Natural Language Processing course. ## Project Overview - **Task:** Abstractive text summarization - **Language:** Somali (low-resource) - **Dataset:** XLSum Somali - **Model:** FLAN-T5-Small - **Fine-tuning:** Low-Rank Adaptation (LoRA) - **Optimization:** INT8 Dynamic Quantization (CPU) - **Evaluation:** ROUGE, BERTScore, latency benchmarking --- ## Repository Structure ├── data/ │ └── raw/ # XLSum Somali JSONL files ├── notebooks/ │ └── main.ipynb # End-to-end experiment notebook ├── src/ │ ├── data/ # Data loading and preprocessing │ ├── models/ # LoRA and quantization logic │ ├── training/ # Training pipeline │ └── demo/ # Gradio demo application ├── results/ │ ├── checkpoints/ # Trained models │ └── metrics/ # Evaluation results (JSON) ├── reports/ │ └── Technical_Report.docx ├── requirements.txt └── README.md --- ## Methodology 1. Load and preprocess Somali BBC news articles. 2. Apply instruction-style prompting for summarization. 3. Fine-tune FLAN-T5 using LoRA adapters. 4. Quantize the trained model using INT8 dynamic quantization. 5. Evaluate quality, latency, and memory trade-offs. 6. Demonstrate inference via a Gradio-based UI. --- ## Results Summary ### FP32 (Merged LoRA) - ROUGE-1: 0.1406 - ROUGE-L: 0.1092 - BERTScore F1: 0.6109 - Latency: 1.83 s/sample ### INT8 Dynamic (CPU) - ROUGE-1: 0.0349 - ROUGE-L: 0.0284 - BERTScore F1: 0.5730 - Latency: 1.60 s/sample - Speed-up: ~1.14× --- ## Demo A Gradio-based UI is included for interactive summarization and comparison between FP32 and INT8 models. --- ## Requirements - Python 3.10+ - PyTorch - Hugging Face Transformers - Datasets - PEFT - Evaluate - Gradio Install dependencies: ```bash pip install -r requirements.txt