Distil-Qwen3-4B-Text2SQL

A fine-tuned Qwen3-4B model for converting natural language questions into SQL queries. Trained using knowledge distillation from DeepSeek-V3, this 4B parameter model matches teacher-level accuracy while being small enough to run locally.

Results

Metric	DeepSeek-V3 (Teacher)	Qwen3-4B (Base)	This Model
LLM-as-a-Judge	80%	62%	80%
Exact Match	48%	16%	60%
ROUGE	87.6%	84.2%	89.5%
METEOR	85.1%	87.3%	86.1%

The fine-tuned model matches the 685B parameter teacher on LLM-as-a-Judge accuracy and exceeds it on exact match and ROUGE scores.

Quick Start

Using Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("distil-labs/distil-qwen3-4b-text2sql")
tokenizer = AutoTokenizer.from_pretrained("distil-labs/distil-qwen3-4b-text2sql")

schema = """CREATE TABLE employees (
  id INTEGER PRIMARY KEY,
  name TEXT NOT NULL,
  department TEXT,
  salary INTEGER
);"""

question = "How many employees earn more than 50000?"

messages = [
    {
        "role": "system",
        "content": """You are a problem solving model working on task_description XML block:
<task_description>You are given a database schema and a natural language question. Generate the SQL query that answers the question.

Input:
- Schema: One or two table definitions in SQL DDL format
- Question: Natural language question about the data

Output:
- A single SQL query that answers the question
- No explanations, comments, or additional text

Rules:
- Use only tables and columns from the provided schema
- Use uppercase SQL keywords (SELECT, FROM, WHERE, etc.)
- Use SQLite-compatible syntax</task_description>
You will be given a single task in the question XML block
Solve only the task in question block.
Generate only the answer, do not generate anything else"""
    },
    {
        "role": "user",
        "content": f"""Now for the real task, solve the task in question block.
Generate only the solution, do not generate anything else
<question>Schema:
{schema}

Question: {question}</question>"""
    }
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Using Ollama (GGUF version)

For local inference, use the quantized GGUF versions:

• distil-qwen3-4b-text2sql-gguf - Full precision GGUF
• distil-qwen3-4b-text2sql-gguf-4bit - 4-bit quantized (~2.5GB)

# Download and create Ollama model
ollama create distil-qwen3-4b-text2sql -f Modelfile

# Run inference
ollama run distil-qwen3-4b-text2sql

Model Details

Property	Value
Base Model	Qwen/Qwen3-4B
Parameters	4 billion
Architecture	Qwen3ForCausalLM
Context Length	262,144 tokens
Precision	bfloat16
Training Data	~10,000 synthetic examples
Teacher Model	DeepSeek-V3

Training

This model was trained using the Distil Labs platform:

1. Seed Data: 50 hand-validated Text2SQL examples covering various SQL complexities
2. Synthetic Generation: Expanded to ~10,000 examples using DeepSeek-V3
3. Fine-tuning: 4 epochs on the synthetic dataset
4. Evaluation: LLM-as-a-Judge with semantic equivalence checking

Training Hyperparameters

• Epochs: 4
• Learning Rate: 5e-5 (cosine schedule)
• Batch Size: 1 (with gradient accumulation)
• Total Steps: ~40,000

Task Format

Input Format

Schema:
CREATE TABLE table_name (
  column_name DATA_TYPE [CONSTRAINTS],
  ...
);

Question: Natural language question about the data

Output Format

A single SQL query with:

• Uppercase SQL keywords (SELECT, FROM, WHERE, etc.)
• SQLite-compatible syntax
• No explanations or additional text

Supported SQL Features

• Simple: SELECT, WHERE, COUNT, SUM, AVG, MAX, MIN
• Medium: JOIN, GROUP BY, HAVING, ORDER BY, LIMIT
• Complex: Subqueries, multiple JOINs, UNION

Use Cases

• Natural language interfaces to databases
• SQL query assistance and autocompletion
• Database chatbots and conversational BI
• Educational tools for learning SQL

Limitations

• Optimized for SQLite syntax
• Best with 1-2 table schemas
• May struggle with highly complex nested subqueries
• Trained on English questions only

License

This model is released under the Apache 2.0 license.

Links

• Distil Labs Website
• GitHub
• Hugging Face

Citation

@misc{distil-qwen3-4b-text2sql,
  author = {Distil Labs},
  title = {Distil-Qwen3-4B-Text2SQL: A Fine-tuned Model for Natural Language to SQL},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/distil-labs/distil-qwen3-4b-text2sql}
}