Neeharika20/qwen-spider-lora

Qwen3 NL → SQL LoRA (Spider-style)

Model Overview

• Developed by: Neeharika20
• Base model: Qwen/Qwen3-4B-Instruct-2507
• Fine-tuning method: LoRA
• Task: Natural Language → SQL generation
• Frameworks: Unsloth, TRL
• License: Apache-2.0

This model is a LoRA adapter fine-tuned to convert natural language questions into SQL queries given an explicit database schema.
It follows the Spider dataset prompt style and is optimized for structured NL → SQL tasks rather than general chat.

Training was accelerated using Unsloth, enabling faster and memory-efficient fine-tuning.

---

Intended Use

✅ Recommended

• Natural Language → SQL generation
• Educational, research, and prototype systems
• Schema-aware SQL generation (Spider-style)

❌ Not Recommended

• General conversational chat
• Schema inference without explicit table definitions
• Production SQL execution without validation

---

Reproducibility

This model can be reproduced by fine-tuning Qwen/Qwen3-4B-Instruct-2507 on the Spider dataset using Hugging Face TRL and Unsloth. Training was performed with a learning rate of 2e-4, max sequence length 2048, 4-bit quantization, and instruction-based prompting including schema context.

Prompt

prompt = """
### Instruction:
Convert the question into an SQL query.

### Database Schema:
CREATE TABLE students(id INT, name TEXT, age INT);

### Question:
List the names of all students.

### SQL:
"""

Limitations

• May hallucinate columns if schema is incomplete
• Performance drops on ambiguous schemas
• SQL syntax is not guaranteed to be valid in all edge cases
• No execution feedback loop during generation

Ethical Considerations

• Trained exclusively on publicly available data
• No personal or sensitive data involved
• Generated SQL should be validated before execution
• Not designed for autonomous decision-making