Datasetpublic

replysadiq/qatar-customs-nl2sql

Qatar Customs NL2SQL — Arabic Natural Language to SAP HANA SQL Fine-tuning dataset for converting Arabic natural language questions into SAP HANA SQL queries for the Qatar General Directorate of Customs (GDC) database. Dataset Overview Metric Value Total question sets 99 Training examples 320 (3 fuzz levels × 80 questions + 80 follow-ups) Test examples 76 (3 fuzz levels × 19 questions + 19 follow-ups) Language Arabic 🇶🇦 (formal + dialectal)… See the full description on the dataset page: https://huggingface.co/datasets/replysadiq/qatar-customs-nl2sql.

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes2downloads

Dataset Card

Qatar Customs NL2SQL — Arabic Natural Language to SAP HANA SQL

Fine-tuning dataset for converting Arabic natural language questions into SAP HANA SQL queries for the Qatar General Directorate of Customs (GDC) database.

Dataset Overview

Metric	Value
Total question sets	99
Training examples	320 (3 fuzz levels × 80 questions + 80 follow-ups)
Test examples	76 (3 fuzz levels × 19 questions + 19 follow-ups)
Language	Arabic 🇶🇦 (formal + dialectal)
SQL dialect	SAP HANA
Domain	Customs declarations, payments, offences, e-declarations
Format	ChatML messages (system/user/assistant)

3 Fuzz Levels

Each question has 3 difficulty variants to train robustness:

Level	Style	Example
Clean	Formal Arabic	خلال آخر 30 يوم، ما إجمالي المبالغ المحصلة عبر الدفع الإلكتروني؟
Medium	Casual dialect	أبغى تحصيلات الدفع الالكتروني آخر 30 يوم
Hard	Very colloquial, abbreviated	كم مجموع التحصيل الكتروني آخر شهر؟

Database Schema (7 Main Tables + 31 Lookup Tables)

Table	Domain
`PBI_VW_FACT_DECLARATION`	Customs declarations
`PBI_VW_FACT_ITEM`	Declaration line items (HS codes, duty)
`PBI_VW_FACT_PARTY`	Importers, exporters, clearing agents
`PBI_VW_FACT_PAYMENT`	Payment/collection records
`PBI_VW_FACT_OFFENCE`	Customs offences/violations
`PBI_VW_FACT_E_DECLARATION`	Traveler e-declarations
`PBI_VW_FACT_OGA`	Other Government Agency checks

Full DDL in eval/schema.sql.

Training

bash

pip install transformers trl torch datasets trackio accelerate peft bitsandbytes

python scripts/train.py

Hardware: Single GPU with 24GB VRAM (A10G, A100, RTX 4090) — QLoRA 4-bit Time: ~2-3 hours Output: LoRA adapter pushed to replysadiq/qatar-customs-nl2sql-qwen7b

Evaluation

bash

python scripts/evaluate.py <model_path>

Metrics: exact match, table accuracy, column accuracy, structural similarity, HANA function accuracy, fuzz degradation.

Files

├── train.jsonl              # 320 training examples
├── test.jsonl               # 76 test examples (held out)
├── eval/
│   ├── test_with_meta.jsonl # Test data with metadata (question_num, fuzz_level)
│   └── schema.sql           # Full HANA DDL for all tables
└── scripts/
    ├── train.py             # QLoRA training script
    ├── evaluate.py          # Evaluation framework
    └── build_dataset.py     # Dataset construction from Excel