Enterprise Text-to-SQL + Voice -- QLoRA fine-tuned Qwen3-8B
Strict execution accuracy: 43.71% to 70.86% on 453 held-out questions.
By
Garvit Audichya - HuggingFace
Features
- Voice input -- speak your question, Whisper transcribes it
- Fine-tuned Qwen3-8B -- QLoRA r=16, 4-bit NF4
- Self-correction loop -- feeds DB errors back to model, retries once
- Static SQL validation -- sqlglot catches bad queries before execution
- PostgreSQL execution -- read-only, statement timeout, no writes possible
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch
BASE = "Qwen/Qwen3-8B"
ADAPTER = "garvit-010/qwen3-8b-text2sql-qlora-v3"
bnb = BitsAndBytesConfig(
load_in_4bit=True, bnb_4bit_quant_type="nf4",
bnb_4bit_use_double_quant=True, bnb_4bit_compute_dtype=torch.float16,
)
base = AutoModelForCausalLM.from_pretrained(BASE, quantization_config=bnb, device_map={"":0})
model = PeftModel.from_pretrained(base, ADAPTER).eval()
tokenizer = AutoTokenizer.from_pretrained(ADAPTER)
IMPORTANT: Always use enable_thinking=False in apply_chat_template.
Results
| Configuration | Strict accuracy |
|---|---|
| Base Qwen3-8B | 43.71% |
| Fine-tuned (this adapter) | 68.43% |
| Fine-tuned + self-correction | 70.86% |
Training
- Base: Qwen/Qwen3-8B @ b968826d9c46, 4-bit NF4
- LoRA r=16, alpha=32, 43.6M trainable params (0.917%)
- Completion-only loss on SQL tokens only
- Kaggle T4, 1 epoch, 135 steps, ~7 hours
License
Apache 2.0 - all training data is synthetic
- Downloads last month
- 55
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support