Views
No views yet
📊 Model Specs
|
🔧 Training Info
|
🎧Customer ServiceAutomated support for e-commerce, SaaS, and service businesses |
💰Sales AssistantsProduct recommendations and lead qualification |
📱WhatsApp BotsConversational AI for messaging platforms |
⚙️AutomationScheduling, FAQs, and workflow automation |
pip install transformers accelerate bitsandbytes torch1from transformers import AutoModelForCausalLM, AutoTokenizer
2
3model = AutoModelForCausalLM.from_pretrained("yoshii-ai/Yoshii-7B-BR", device_map="auto")
4tokenizer = AutoTokenizer.from_pretrained("yoshii-ai/Yoshii-7B-BR")
5
6messages = [{"role": "user", "content": "Ola, preciso de ajuda com meu pedido"}]
7inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
8outputs = model.generate(inputs, max_new_tokens=256, do_sample=True, temperature=0.7)
9print(tokenizer.decode(outputs[0], skip_special_tokens=True))1import torch
2from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
3
4bnb_config = BitsAndBytesConfig(
5 load_in_4bit=True,
6 bnb_4bit_quant_type="nf4",
7 bnb_4bit_compute_dtype=torch.float16,
8 bnb_4bit_use_double_quant=True,
9)
10
11model = AutoModelForCausalLM.from_pretrained(
12 "yoshii-ai/Yoshii-7B-BR",
13 quantization_config=bnb_config,
14 device_map="auto",
15)
16tokenizer = AutoTokenizer.from_pretrained("yoshii-ai/Yoshii-7B-BR")
17
18# Your conversation here
19messages = [{"role": "user", "content": "Qual o horario de funcionamento?"}]
20inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")
21outputs = model.generate(inputs, max_new_tokens=256, do_sample=True, temperature=0.7)
22print(tokenizer.decode(outputs[0], skip_special_tokens=True))|
🎧 Customer Support
|
📦 Product Inquiry
|
|
🚚 Shipping Info
|
🕐 Business Hours
|
1# QLoRA Configuration
2LoraConfig(
3 r=32, # LoRA rank
4 lora_alpha=64, # LoRA alpha
5 lora_dropout=0.05, # Dropout
6 bias="none",
7 task_type="CAUSAL_LM",
8 target_modules=["q_proj", "v_proj"]
9)
10
11# Training Arguments
12SFTConfig(
13 num_train_epochs=3,
14 per_device_train_batch_size=1,
15 gradient_accumulation_steps=16,
16 learning_rate=2e-4,
17 lr_scheduler_type="cosine",
18 warmup_ratio=0.03,
19 max_length=512,
20 fp16=True,
21 gradient_checkpointing=True,
22 optim="paged_adamw_8bit",
23)
24
25# Quantization
26BitsAndBytesConfig(
27 load_in_4bit=True,
28 bnb_4bit_quant_type="nf4",
29 bnb_4bit_compute_dtype=torch.float16,
30 bnb_4bit_use_double_quant=True,
31)| Metric | Value |
|---|---|
| Training Loss (final) | ~0.8 |
| Inference Speed | ~30 tokens/sec (RTX 3070) |
| Memory Usage | ~5.8 GB VRAM |
1@misc{yoshii7bbr2025,
2 author = {Richard Sakaguchi},
3 title = {Yoshii-7B-BR: Brazilian Portuguese Customer Service LLM},
4 year = {2025},
5 publisher = {Hugging Face},
6 howpublished = {\url{https://huggingface.co/yoshii-ai/Yoshii-7B-BR}},
7}
