qwen2.5-coder-3b-sql-doctor-mlx

A domain-adapted Small Language Model (SLM) for offline SQL performance diagnosis, query rewriting, and index recommendation on Apple Silicon.

Fine-tuned from mlx-community/Qwen2.5-Coder-3B-Instruct-4bit by Kartik Raktate.

---

🎯 Motivation: Privacy & Zero Cloud Dependency

Enterprise production databases contain sensitive customer schemas, confidential column names, and proprietary business logic. Sending production queries to third-party cloud LLMs creates compliance risks and ongoing API costs.

qwen2.5-coder-3b-sql-doctor-mlx runs 100% locally and offline on Apple Silicon via Apple's native MLX framework. It acts as an on-device Senior Database Administrator (DBA) for developer CLI tools, local IDEs, and offline CI/CD pipelines.

---

⚡ Key Highlights

• Edge Optimized: Quantized to 4-bit group-wise INT4 precision. Runs with only ~2.1 GB of RAM on Apple Silicon unified memory.
• Strict Domain Output: Generates clean, structured technical reports with:
  1. Root-Cause Diagnosis: Exact explanation of why the query suffers from sequential scans, memory spills, or CPU bottlenecks.
  2. Optimized SQL: Rewritten query eliminating anti-patterns (SELECT *, unindexed subqueries, functions on indexed columns).
  3. Migration Index DDL: Ready-to-run CREATE INDEX CONCURRENTLY ... statement for PostgreSQL or MySQL.
• Sub-Second Latency: Delivers instantaneous token generation locally using Metal GPU acceleration on Mac.

---

🚀 Quickstart: Running Locally with MLX

1. Installation

pip install mlx-lm

2. Python Inference

from mlx_lm import load, generate

# Load model directly from Hugging Face Hub
model_id = "KarRak007/qwen2.5-coder-3b-sql-doctor-mlx"
model, tokenizer = load(model_id)

# Prompt with your slow query and table schema
user_prompt = """Query: SELECT * FROM orders WHERE YEAR(order_date) = 2024;
Schema: orders (id BIGINT, user_id BIGINT, total NUMERIC, status VARCHAR, order_date DATE)"""

messages = [
    {
        "role": "system",
        "content": "You are sql-doctor, an expert Database Administrator and SQL performance engineer. Analyze the query and provide: 1) Diagnosis, 2) Optimized Query, and 3) Recommended Index."
    },
    {"role": "user", "content": user_prompt}
]

prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=prompt, max_tokens=350, verbose=False)
print(response)

3. Interactive Terminal Chat

You can test the model interactively directly in your terminal:

python3 -m mlx_lm.chat --model KarRak007/qwen2.5-coder-3b-sql-doctor-mlx

---

💡 Example Evaluation

Input Prompt:

Query: SELECT * FROM audit_logs WHERE action LIKE '%login_failed%' ORDER BY created_at DESC LIMIT 50;
Schema: audit_logs (id BIGINT, action VARCHAR, entity_type VARCHAR, entity_id BIGINT, created_at TIMESTAMP)

Model Output:

### Diagnosis
1. `LIKE '%pattern%'` with a leading wildcard cannot leverage a standard B-Tree index, forcing an expensive full sequential scan over disk.
2. `SELECT *` incurs unnecessary I/O overhead reading unused column blocks.

### Optimized Query
```sql
SELECT id, action, entity_id, created_at 
FROM audit_logs 
WHERE action LIKE '%login_failed%' 
ORDER BY created_at DESC 
LIMIT 50;

Recommended Index

CREATE EXTENSION IF NOT EXISTS pg_trgm;
CREATE INDEX CONCURRENTLY idx_audit_logs_action_gin ON audit_logs USING gin (action gin_trgm_ops);

---

## 🛠️ Companion Tool: `sql-doctor` CLI

This model powers **`sql-doctor`**, an open-source hybrid static AST analyzer and database query doctor that runs locally in terminal workflows.

```bash
# Check query anti-patterns locally
sql-doctor check "SELECT * FROM users JOIN orders WHERE email LIKE '%@gmail.com'"

---

📊 Technical Specifications

Parameter	Specification
Base Architecture	Qwen 2.5 Coder 3B Instruct
Quantization	4-bit INT4 Group-wise
Trained Layers	LoRA Attention & MLP Adaptors ($r=8, \alpha=16$)
Context Window	32,768 tokens
Memory Footprint	~2.1 GB VRAM / Unified Memory
Inference Framework	Apple MLX (mlx-lm)
Primary Dialects	PostgreSQL, MySQL, ANSI SQL

---

📄 License & Attribution

This model is licensed under the Apache 2.0 License.
Developed and maintained by Kartik Raktate.