SoraniSQL: A Novel Synthetic Text-to-SQL Dataset and Benchmark for Low-Resource Central Kurdish (Sorani)
Year: 2026
Type: preprint
Abstract: Text-to-SQL systems allow users to query databases in natural language, yet existing benchmarks cover English and a small number of well-resourced languages, and no resource has been reported for Kurdish. This paper presents SoraniSQL, a novel dataset and benchmark that is, to the best of current knowledge, the first for Text-to-SQL in Central Kurdish (Sorani). The dataset contains 7,518 question-SQL pairs from 609 query types, systematically constructed over a nine-table retail point-of-sale database with four database instances for test-suite execution accuracy (EXₜₛ); a five-member native-speaker panel rated a stratified sample of 200 questions at 4.9 out of 5 for fluency and 4.9 for adequacy. Dedicated splits separate new values, unseen phrasings, unseen combinations of query components, unseen SQL constructs and spelling variation. Experiments with retrieval, template, compositional, neural and pretrained models, all tuned on development data, show that a template classifier reaches 97.1% EXₜₛ on new values but 0.7% on unseen combinations, where a compositional model reaches 97.7%. A 1.5B-parameter code model fine-tuned with quantised low-rank adaptation (QLoRA) reaches 95.0% on unseen combinations on a laptop central processing unit (CPU), and the structure split and the orthographic-noise sets form the most demanding conditions of the benchmark. A Windows application, DanaText2SQL, demonstrates offline use of the benchmark's best fast model with Microsoft SQL Server.
Source: Research Square
Author: Dana Lattef Hussein¹
Institution: Sulaimani Polytechnic University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7226014860
