Lloro SQL, developed by Semantix Research Labs, is a language Model that was trained to effectively transform Portuguese queries into SQL Code. It is a fine-tuned version of meta-llama/Meta-Llama-3-8B-Instruct, that was trained on GretelAI public datasets. The fine-tuning process was performed using the QLORA metodology on a GPU A100 with 40 GB of RAM.

Model description

Model type: A 7B parameter fine-tuned on GretelAI public datasets.

Language(s) (NLP): Primarily Portuguese, but the model is capable to understand English as well

Finetuned from model: meta-llama/Meta-Llama-3-8B-Instruct

What is Lloro's intended use(s)?

Lloro is built for Text2SQL in Portuguese contexts .

Input : Text

Output : Text (Code)

Usage

Using an OpenAI compatible inference server (like vLLM)

python

from openai import OpenAI
client = OpenAI(
    api_key="EMPTY",
    base_url="http://localhost:8000/v1",
)
def generate_responses(instruction, client=client):
    
    chat_response = client.chat.completions.create(
    model=<model>,
    messages=[
        {"role": "system", "content": "Você escreve a instrução SQL que responde às perguntas feitas. Você NÃO FORNECE NENHUM COMENTÁRIO OU EXPLICAÇÃO sobre o que o código faz, apenas a instrução SQL terminando em ponto e vírgula. Você utiliza todos os comandos disponíveis na especificação SQL, como: [SELECT, WHERE, ORDER, LIMIT, CAST, AS, JOIN]."},
        {"role": "user", "content": instruction},
    ]
)
    
    return chat_response.choices[0].message.content

output = generate_responses(user_prompt)

Params

Training Parameters

Params	Training Data	Examples	Tokens	LR
8B	GretelAI public datasets + Synthetic Data	102970	18.654.222	2e-4

Model Sources

GretelAI: https://huggingface.co/datasets/gretelai/synthetic_text_to_sql

Performance

Test Dataset

Model	LLM as Judge	Code Bleu Score	Rouge-L	CodeBert- Precision	CodeBert-Recall	CodeBert-F1	CodeBert-F3
Llama 3 8B	65.48%	0.4583	0.6361	0.8815	0.8871	0.8835	0.8862
Lloro - SQL	71.33%	0.6512	0.7965	0.9458	0.9469	0.9459	0.9466
GPT - 3.5 Turbo	67.52%	0.6232	0.9967	0.9151	0.9152	0.9142	0.9175

Database Benchmark

Model	Score
Llama 3 - Base	35.55%
Lloro - SQL	49.48%
GPT - 3.5 Turbo	46.15%

Translated BIRD Benchmark - https://bird-bench.github.io/

Model	Score
Llama 3 - Base	33.87%
Lloro - SQL	47.14%
GPT - 3.5 Turbo	42.14%

Training Infos

The following hyperparameters were used during training:

Parameter	Value
learning_rate	2e-4
weight_decay	0.001
train_batch_size	16
eval_batch_size	8
seed	42
optimizer	Adam - adamw_8bit
lr_scheduler_type	cosine
num_epochs	4.0

QLoRA hyperparameters

The following parameters related with the Quantized Low-Rank Adaptation and Quantization were used during training:

Parameter	Value
lora_r	64
lora_alpha	128
lora_dropout	0

Experiments

Model	Epochs	Overfitting	Final Epochs	Training Hours	CO2 Emission (Kg)
Llama 3 8B Instruct	5	Yes	4	10.16	1.45

Framework versions

Library	Version
accelerate	0.21.0
bitsandbytes	0.42.0
Datasets	2.14.3
peft	0.4.0
Pytorch	2.0.1
safetensors	0.4.1
scikit-image	0.22.0
scikit-learn	1.3.2
Tokenizers	0.14.1
Transformers	4.37.2
trl	0.4.7

semantixai/Lloro-SQL

Lloro SQL