dmatekenya/chichewa-text2sql

Text-to-SQL system for Chichewa, a low-resource African language. The project explores LLM-based and hybrid approaches for translating natural language questions into SQL queries, supporting database-driven analytics and chatbot applications in low-resource settings. # Chichewa Text-to-SQL > **The first Text-to-SQL benchmark for Chichewa** — a low-resource Bantu language spoken by over 12 million people in Malawi and neighboring regions. ## Overview Recent advances in Large Language Models (LLMs) have significantly improved Text-to-SQL performance in high-resource languages. However, their effectiveness in low-resource language settings remains largely underexplored. This work investigates the adaptation of LLMs for Text-to-SQL generation in Chichewa. We construct a structured Chichewa Text-to-SQL benchmark consisting of **400 manually curated natural language–SQL pairs** grounded in a unified relational database covering agriculture, commodity prices, population statistics, market data, and food insecurity. We systematically evaluate open-source LLMs under zero-shot, random few-shot, and retrieval-augmented few-shot prompting in both **English and Chichewa**, and fine-tune the best-performing model using QLoRA. **Keywords:** Text-to-SQL · Low-Resource Languages · Chichewa · QLoRA Fine-Tuning · Semantic Parsing · Information Retrieval --- ## Dataset The benchmark contains **400 manually curated natural language–SQL pairs** across 5 database tables, split into train (280) / dev (60) / test (60): | Table | Description | Examples | |---|---|---| | `commodity_prices` | Market prices for 6 crops across 27 districts (2024) | 80 | | `production` | Crop yields for 46 crops across 28 districts (2023–2024) | 80 | | `population` | District-level population statistics | 80 | | `food_insecurity` | Food insecurity indicators by district | 80 | | `mse_daily` | Malawi Stock Exchange daily market data | 80 | Each example includes: - `question_ny` — question in Chichewa (Nyanja) - `question_en` — question in English - `sql_statement` — ground-truth SQL query - `table` — target database table **Data splits:** `train.json` · `dev.json` · `test.json` · `all.json` > The raw data files are not versioned here (see `.gitignore`). The SQ …