qwen3-1.7b-vivu-travel-vn is a 1.7 Billion parameter Small Language Model (SLM) domain-specifically fine-tuned for the Vietnamese Tourism Industry. Built upon the advanced architecture of Qwen3-1.7B, this model serves as ViVu, an intelligent virtual assistant specifically designed for seamless integration into Retrieval-Augmented Generation (RAG) pipelines.

2. Core Capabilities

Strict Anti-Hallucination: Underwent rigorous instruction-tuning with a zero-tolerance policy for fabricated information. Responses are 100% grounded in the provided input context. The model automatically and politely declines to answer if the context lacks relevant data.
RAG-Optimized: Demonstrates excellent comprehension and synthesis of information chunks retrieved from Vector Databases, triggered by XML-style tags (<context>...</context>).
Disabled Thinking Mode (Low Latency): Unlike the base Qwen3 model, this fine-tuned version is designed to operate with Thinking Mode disabled. This forces the model to act as a direct, fast RAG extractor rather than a multi-step reasoner, drastically reducing latency and preventing infinite reasoning loops in real-time chat applications.
Edge-Deployable: With its compact 1.7B footprint, the model requires minimal VRAM, allowing for deployment on consumer-grade GPUs (e.g., RTX 3060, T4) or local execution on CPUs/Mobile NPUs via GGUF format.

3. Model Specification

Base Model: unsloth/Qwen3-1.7B
Parameters: 1.7 Billion
Architecture: Causal Language Modeling (Transformer), 32k context length.
Training Methodology: LoRA Instruction-tuning via Unsloth (train_on_responses_only).
Thinking Mode: Disabled (Standard LLM behavior).
Supported Languages: Vietnamese (Primary), English.

4. Training & Evaluation Data

The model's performance is driven by a high-quality, manually curated and labeled dataset, divided into two primary focuses:

Training Dataset: Comprises ~2,500 diverse samples covering the comprehensive Vietnamese tourism ecosystem (locations, cuisine, itineraries, pricing, local culture). The data is strictly formatted to RAG standards (Context - Query - Response) to condition the model for precise information extraction rather than free-form text generation.
Multi-Reference Test Set: To ensure objective academic evaluation, the Test Set is designed with a Multi-Reference structure. Each query contains multiple Ground Truth answers, representing different natural language styles while preserving absolute factual accuracy. This reflects the diversity of natural language and ensures fair, robust scoring when calculating automated metrics (ROUGE, BLEU) and conducting Blind Human Evaluations.

python

# pip install transformers accelerate

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "thanhdo881/qwen3-1.7b-vivu-travel-vn" 

# 1. Load Model & Tokenizer
model = AutoModelForCausalLM.from_pretrained(
    model_name, 
    torch_dtype=torch.bfloat16, # Or float16 depending on your hardware
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# 2. Strict System Prompt (Mandatory to trigger the ViVu persona)
# NOTE: Keep this in Vietnamese as the model was fine-tuned on this exact string.
SYSTEM_PROMPT = "Bạn là trợ lý du lịch AI chuyên nghiệp. Hãy trả lời chính xác dựa trên ngữ cảnh được cung cấp. Yêu cầu bắt buộc: Chỉ sử dụng 100% TIẾNG VIỆT, tuyệt đối không sử dụng tiếng Trung, tiếng Anh hay bất kỳ ngôn ngữ nào khác."

# 3. Prepare Input (Using the standard XML tag structure)
context = "Đà Lạt nằm trên cao nguyên Lâm Viên, nổi tiếng với khí hậu ôn đới và Hồ Xuân Hương."
question = "Đà Lạt có những đặc điểm gì nổi bật?"

user_content = f"<context>\n{context}\n</context>\n\nCÂU HỎI:\n{question}"

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": user_content}
]

# 4. Generate Text (CRITICAL: Disable Thinking Mode)
text = tokenizer.apply_chat_template(
    messages, 
    tokenize=False, 
    add_generation_prompt=True,
    enable_thinking=False # Must be False for this fine-tune
)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Recommended parameters for Non-Thinking Qwen3
outputs = model.generate(
    **inputs, 
    max_new_tokens=256, 
    temperature=0.7, 
    top_p=0.8,
    do_sample=True
)
response = tokenizer.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0]

print(response)