Lloro, developed by Semantix Research Labs , is a language Model that was trained to effectively perform Portuguese Data Analysis in Python. It is a fine-tuned version of codellama/CodeLlama-7b-Instruct-hf, that was trained on synthetic 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 synthetic datasets.

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

Finetuned from model: codellama/CodeLlama-7b-Instruct-hf

What is Lloro's intended use(s)?

Lloro is built for data analysis in Portuguese contexts .

Input : Text

Output : Text (Code)

Usage

Using Transformers

python

#Import required libraries
import torch
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer
)

#Load Model
model_name = "semantixai/LloroV2"
base_model = AutoModelForCausalLM.from_pretrained(
        model_name,
        return_dict=True,
        torch_dtype=torch.float16,
        device_map="auto",
    )

#Load Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)


#Define Prompt
user_prompt = "Desenvolva um algoritmo em Python para calcular a média e a mediana dos preços de vendas por tipo de material do produto."
system = "Provide answers in Python without explanations, only the code"
prompt_template = f"[INST] <<SYS>>\\n{system}\\n<</SYS>>\\n\\n{user_prompt}[/INST]"

#Call the model
input_ids = tokenizer([prompt_template], return_tensors="pt")["input_ids"].to("cuda")

            
outputs = base_model.generate(
    input_ids,
    do_sample=True,
    top_p=0.95,
    max_new_tokens=1024,
    temperature=0.1,
    )

#Decode and retrieve Output
output_text = tokenizer.batch_decode(outputs, skip_prompt=True, skip_special_tokens=False)
display(output_text)

Using an OpenAI compatible inference server (like vLLM)

python

from openai import OpenAI

client = OpenAI(
    api_key="EMPTY",
    base_url="http://localhost:8000/v1",
)
user_prompt = "Desenvolva um algoritmo em Python para calcular a média e a mediana dos preços de vendas por tipo de material do produto."
completion = client.chat.completions.create(temperature=0.1,frequency_penalty=0.1,model="semantixai/Lloro",messages=[{"role":"system","content":"Provide answers in Python without explanations, only the code"},{"role":"user","content":user_prompt}])

Params Training Parameters

Params	Training Data	Examples	Tokens	LR
7B	Pairs synthetic instructions/code	74222	9 351 532	2e-4

Model Sources

Test Dataset Repository: https://huggingface.co/datasets/semantixai/LloroV3

Model Dates: Lloro was trained between February 2024 and April 2024.

Performance

Modelo	LLM as Judge	Code Bleu Score	Rouge-L	CodeBert- Precision	CodeBert-Recall	CodeBert-F1	CodeBert-F3
GPT 3.5	94.29%	0.3538	0.3756	0.8099	0.8176	0.8128	0.8164
Instruct -Base	88.77%	0.3666	0.3351	0.8244	0.8025	0.8121	0.8052
Instruct -FT	97.95%	0.5967	0.6717	0.9090	0.9182	0.9131	0.9171

Training Infos: The following hyperparameters were used during training:

Parameter	Value
learning_rate	2e-4
weight_decay	0.0001
train_batch_size	7
eval_batch_size	7
seed	42
optimizer	Adam - paged_adamw_32bit
lr_scheduler_type	cosine
lr_scheduler_warmup_ratio	0.06
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	256
lora_dropout	0.1
storage_dtype	"nf4"
compute_dtype	"bfloat16"

Experiments

Model	Epochs	Overfitting	Final Epochs	Training Hours	CO2 Emission (Kg)
Code Llama Instruct	1	No	1	3.01	0.43
Code Llama Instruct	4	Yes	3	9.25	1.32

Framework versions

Library	Version
bitsandbytes	0.40.2
Datasets	2.14.3
Pytorch	2.0.1
Tokenizers	0.14.1
Transformers	4.34.0