visheshgupta/mistral-7b-text2sql-qlora

Mistral-7B Text-to-SQL (QLoRA)

A QLoRA fine-tuned adapter for Mistral-7B-v0.3 that converts natural language questions into SQL queries.

Trained on the sql-create-context dataset (WikiSQL + Spider, 20K examples) using a single Kaggle T4 GPU in 4h 45min.

Results

Metric	Base Mistral-7B	This Adapter (v3)	Delta
Execution Accuracy	57.5%	94.0%	+36.5pp
Exact Match	1.0%	74.0%	+73pp
Avg BLEU	0.428	0.923	+0.495
Valid SQL Rate	100.0%	100.0%	±0

Evaluated on 200 samples with SQL executed against SQLite.

Usage

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

base_model = "mistralai/Mistral-7B-v0.3"
adapter = "visheshgupta29/mistral-7b-text2sql-qlora"

# Load quantized base model
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True,
)
model = AutoModelForCausalLM.from_pretrained(
    base_model, quantization_config=bnb_config, device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter)
model.eval()

tokenizer = AutoTokenizer.from_pretrained(adapter)

# Build prompt
prompt = """### Task: Generate a single SQL query to answer the following question.
Do not repeat any conditions or output multiple queries.

### Database Schema:
CREATE TABLE employees (id INT, name TEXT, department TEXT, salary REAL);

### Question:
What is the average salary per department?

### SQL Query:
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

# Find all ';' token variants for stop condition
semicolon_ids = set(tokenizer.encode(";", add_special_tokens=False))
for tok_str, tok_id in tokenizer.get_vocab().items():
    if tok_str.replace("\u2581", "").strip() == ";":
        semicolon_ids.add(tok_id)
stop_ids = list({tokenizer.eos_token_id} | semicolon_ids)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=128,
        do_sample=False,
        repetition_penalty=1.1,
        eos_token_id=stop_ids,
    )

generated = output[0][inputs["input_ids"].shape[1]:]
sql = tokenizer.decode(generated, skip_special_tokens=True).strip()
if ";" in sql:
    sql = sql.split(";")[0].strip()
print(sql)
# SELECT department, AVG(salary) FROM employees GROUP BY department

Or use the project's SQLPredictor class:

from src.inference.predict import SQLPredictor

predictor = SQLPredictor(adapter_path="visheshgupta29/mistral-7b-text2sql-qlora")
sql = predictor.predict(
    question="What is the average salary per department?",
    schema="CREATE TABLE employees (id INT, name TEXT, department TEXT, salary REAL);"
)

Training Details

Parameter	Value
Base model	Mistral-7B-v0.3
Method	QLoRA (4-bit NF4 + LoRA)
LoRA rank	16
LoRA alpha	32
Target modules	q, k, v, o, gate, up, down proj
Trainable params	42M (1.1% of 3.8B)
Dataset	sql-create-context (20K train)
Epochs	1
Batch size	2 (effective 16 via grad accum)
Optimizer	Paged AdamW 8-bit
LR	2e-4, cosine schedule
Train loss	0.038
Eval loss	0.025
Training time	4h 45min on Kaggle T4
Peak VRAM	5.37 GB

Key Design Decisions

1. Semicolon in training data — Every SQL completion ends with ; so the model learns to stop generating. Without this, the model repeats conditions endlessly.
2. SentencePiece vocab scan — tokenizer.encode(";") returns ▁; but the model generates bare ; (different token ID). We scan get_vocab() for all variants.
3. repetition_penalty=1.1 — Higher values (1.3, 1.5) cause gibberish on SQL because they penalize tokens from the prompt that must be reused.
4. No no_repeat_ngram_size — SQL has valid repeated n-grams like = "val" AND that this constraint blocks.

Full Project

See the GitHub repository for the complete pipeline: data prep, training, evaluation, comparison, and Gradio demo.

License

MIT