Model Card for 522H0134-NguyenNhatHuy/Vinallama-2.7b-chat-SFT

This model is a fine-tuned version of viet-mistral/vinallama-2.7b-chat using LoRA + PEFT, targeting Vietnamese open-domain, instruction-following chat. It is aligned for safe, helpful, and fluent conversations in Vietnamese through supervised fine-tuning on high-quality prompt-response pairs.

🧠 Model Details

Base Model: viet-mistral/vinallama-2.7b-chat
Model Type: Causal Language Model (Chat)
Languages: Vietnamese
License: Apache 2.0
Fine-tuning Framework: PEFT with LoRA
Training Dataset: Custom Vietnamese SFT & DPO dataset (~10K SFT + 10K DPO + 1K test prompts)

✅ Intended Uses

Direct Use

Vietnamese open-domain dialogue
Instruction-following tasks
Educational or research-based QA

Out-of-Scope Use

Medical, legal, or financial advice
Content moderation or safety-critical tasks
English-centric prompts

🧪 Evaluation

Test Data

The model was evaluated on a Vietnamese test set of 1,000 prompts (60% safe / 40% adversarial) adapted from JailBreak, HarmBench, and OpenAssistant.

Metrics

Helpfulness
Toxicity (via Detoxify > 0.5)
Appropriateness / Safety Rejection

Detoxify was used to filter harmful generations during evaluation.

Summary

74% of generations were rated safe/aligned
86% rejection rate on highly toxic or adversarial prompts
The model avoids unsafe completions better than its base model

🚀 How to Use the Model

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# Load base model and LoRA adapter
tokenizer = AutoTokenizer.from_pretrained("viet-mistral/vinallama-2.7b-chat")
base_model = AutoModelForCausalLM.from_pretrained("viet-mistral/vinallama-2.7b-chat")
model = PeftModel.from_pretrained(base_model, "522H0134-NguyenNhatHuy/Vinallama-2.7b-chat-SFT")

# Chat example
prompt = "Xin chào, bạn có thể giúp tôi học tiếng Anh không?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=150)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))