✨ YOSHII-7B-BR ✨

Yoshii é uma LLM de última geração em Português Brasileiro, especializada em atendimento ao cliente e automação empresarial.

Construído com QLoRA sobre Mistral 7B, oferecendo performance enterprise com requisitos de hardware acessíveis.

---

🎯 Overview

📊 Model Specs Attribute Value Architecture Mistral 7B Parameters 7.2B Context Window 32,768 tokens Quantization 4-bit NF4 Model Size 4.7 GB VRAM Required ~6 GB

🔧 Training Info Attribute Value Method QLoRA Dataset 755 conversations Epochs 3 LoRA Rank 32 Training Time 37 min Hardware RTX 3070 8GB

---

💼 Use Cases

🎧 Customer Service Automated support for e-commerce, SaaS, and service businesses

💰 Sales Assistants Product recommendations and lead qualification

📱 WhatsApp Bots Conversational AI for messaging platforms

⚙️ Automation Scheduling, FAQs, and workflow automation

---

🚀 Quick Start

Installation

pip install transformers accelerate bitsandbytes torch

Basic Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("yoshii-ai/Yoshii-7B-BR", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("yoshii-ai/Yoshii-7B-BR")

messages = [{"role": "user", "content": "Ola, preciso de ajuda com meu pedido"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=256, do_sample=True, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Optimized Usage (4-bit)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

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(
    "yoshii-ai/Yoshii-7B-BR",
    quantization_config=bnb_config,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("yoshii-ai/Yoshii-7B-BR")

# Your conversation here
messages = [{"role": "user", "content": "Qual o horario de funcionamento?"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")
outputs = model.generate(inputs, max_new_tokens=256, do_sample=True, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

💬 Example Conversations

🎧 Customer Support User: Ola, gostaria de cancelar meu pedido Yoshii: Ola! Claro, posso ajudar com o cancelamento. Poderia informar o numero do pedido?	📦 Product Inquiry User: Esse produto tem garantia? Yoshii: Sim! Todos os nossos produtos possuem garantia de 12 meses contra defeitos.
🚚 Shipping Info User: Quanto custa o frete para SP? Yoshii: O frete para Sao Paulo varia conforme o CEP. Pode informar seu CEP?	🕐 Business Hours User: Qual o horario de atendimento? Yoshii: Nosso atendimento funciona de segunda a sexta, das 8h as 18h.

---

📈 Training Details

🔧 Full Training Configuration

# QLoRA Configuration
LoraConfig(
    r=32,                              # LoRA rank
    lora_alpha=64,                     # LoRA alpha
    lora_dropout=0.05,                 # Dropout
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=["q_proj", "v_proj"]
)

# Training Arguments
SFTConfig(
    num_train_epochs=3,
    per_device_train_batch_size=1,
    gradient_accumulation_steps=16,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    max_length=512,
    fp16=True,
    gradient_checkpointing=True,
    optim="paged_adamw_8bit",
)

# Quantization
BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True,
)

---

📊 Performance

Metric	Value
Training Loss (final)	~0.8
Inference Speed	~30 tokens/sec (RTX 3070)
Memory Usage	~5.8 GB VRAM

---

⚠️ Limitations

• Language: Optimized for Brazilian Portuguese only
• Domain: Best suited for customer service contexts
• Responses: May require fine-tuning for specialized domains
• Hallucinations: Like all LLMs, may generate inaccurate information

---

📝 Citation

@misc{yoshii7bbr2025,
  author       = {Richard Sakaguchi},
  title        = {Yoshii-7B-BR: Brazilian Portuguese Customer Service LLM},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/yoshii-ai/Yoshii-7B-BR}},
}

---

🛠️ Built with

🤗 Yoshii AI • 🌐 Sakaguchi IA • 📊 Dataset

---

_{In Memoriam — Yoshii ✦}