tanvicas/nano-analyst-sql

🤖 Nano-Analyst: Fine-Tuned SQL Agent

A private, on-device SQL generation model fine-tuned on the Spider dataset.

Model Description

This is a LoRA fine-tuned version of Llama-3-8B-Instruct specialized for Text-to-SQL generation. The model was trained on 6,300 examples from the Spider dataset and achieves 100% valid SQL generation.

Key Features

• ✅ Privacy-Preserving: Runs locally, no API calls required
• ✅ Efficient: QLoRA fine-tuning (only 1.03% parameters trained)
• ✅ Production-Ready: Includes self-correction and RAG retrieval
• ✅ Well-Documented: Complete training and evaluation pipeline

Model Details

• Base Model: meta-llama/Meta-Llama-3-8B-Instruct
• Fine-Tuning Method: QLoRA (4-bit quantization)
• Framework: Unsloth (2x faster training)
• Training Data: Spider dataset (6,300 training examples)
• LoRA Rank: 32
• LoRA Alpha: 64

Training Results

• Final Training Loss: 0.19
• Validation Loss: 0.31
• Training Steps: 1,182
• Training Time: ~3 hours on T4 GPU

Evaluation Results

Evaluated on 100 examples from Spider validation set:

Metric	Value
Valid SQL Generation	100.0%
Exact String Match	2.0%
Successful Queries	100/100

Usage

Loading the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Meta-Llama-3-8B-Instruct",
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load LoRA adapters
model = PeftModel.from_pretrained(
    base_model,
    "tanvicas/nano-analyst-sql"
)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct")

Generating SQL

question = "How many users are from California?"
schema = "CREATE TABLE users (id INT, name TEXT, state TEXT);"

# Generate SQL using the model
sql = generate_sql(question, schema)
print(sql)  # SELECT COUNT(*) FROM users WHERE state = 'California';

License

Apache 2.0 License. Base model subject to Meta's Llama 3 Community License.

Contact

• GitHub: github.com/tanvicas/nano-analyst

---

Built for learning and research 🚀