salmane11/Darija-to-SQL

Darija2SQL-3B: Darija to SQL Code LLM

Darija2SQL-3B is a specialized 3.1 billion parameter Code LLM developed by salmane11, built upon the Qwen2.5-coder-3B architecture. Its primary function is to translate natural language questions posed in Moroccan Arabic (Darija) into executable SQL queries.

Key Capabilities & Features

• Darija-to-SQL Translation: Directly converts Darija natural language into SQL, enabling database interaction for dialectal Arabic speakers.
• Specialized Fine-tuning: Utilizes Parameter-Efficient Fine-Tuning (PEFT) with LoRA adapters on the Dialect2SQL dataset to maintain strong code reasoning while adapting to Darija.
• Schema Comprehension: Incorporates training for understanding database schemas and normalizing dialectal variations to infer SQL intent.
• Base Model: Inherits robust code reasoning capabilities from the Qwen2.5-coder-3B model.

Intended Use Cases

This model is designed for:

• Research and Education: Particularly in Natural Language to SQL generation for dialectal Arabic.
• Low-Resource Dialect Adaptation: Demonstrates how code models can be adapted for specific, less-resourced dialects.
• Arabic Database Interfaces: Facilitating user interaction with databases using native Arabic dialects.

While effective on its training data, the model's generalization to entirely unseen schemas or domains may require further adaptation. It offers a unique solution for bridging linguistic and technical gaps in database querying.