qwen3-4b-vihsd-lora

Qwen3-4B ViHSD – SFT LoRA Adapter

This repository contains the Supervised Fine-Tuning (SFT) LoRA adapter for Vietnamese hate speech detection with explanations, trained on the ViHSD Explainable dataset.

⚠️ Important:
This repo ONLY contains the LoRA adapter weights, NOT a full merged model.

For the fully merged model (ready for inference), see:
👉 https://huggingface.co/vominhmanh/qwen3-4b-vihsd

---

🔍 Task

Vietnamese content moderation with explainability:

• Classification: CLEAN | OFFENSIVE | HATE
• Explanation: Short Vietnamese rationale
• Evidence: Verbatim spans (if applicable)

Output format (JSON):

{
  "explanation": "...",
  "evidence": ["..."],
  "label": "CLEAN | OFFENSIVE | HATE"
}

---

🧠 Base Model

• Base model: unsloth/Qwen3-4B-Instruct-2507
• Training method: Parameter-Efficient Fine-Tuning (LoRA)
• Framework: Unsloth + TRL (SFTTrainer)

---

📊 Training Data

• Dataset: vominhmanh/vihsd-explainable
• Derived from: htdung167/ViHSD
• Splits: train / validation / test
• Language: Vietnamese

Each sample contains:

• text
• label
• explanation
• evidence

---

🏋️ Training Details

• Method: Supervised Fine-Tuning (SFT)
• Objective: Cross-entropy (NLL)
• LoRA rank: r = 16
• LoRA alpha: 32
• Max sequence length: 512
• Precision: bfloat16
• Quantization: 4-bit base model

This SFT stage focuses on:

• Correct label prediction
• Strict JSON formatting
• Faithful explanation & evidence generation

---

📈 Evaluation

Evaluation was performed on the ViHSD test split (3-class classification):

Note: Explanation and evidence quality were evaluated qualitatively. Preference optimization is handled in a subsequent DPO stage.

---

🔌 How to Use (Load LoRA)

from unsloth import FastLanguageModel
from peft import PeftModel
import torch

# Load base model
model, tokenizer = FastLanguageModel.from_pretrained(
    "unsloth/Qwen3-4B-Instruct-2507",
    max_seq_length=512,
    dtype=torch.bfloat16,
    load_in_4bit=True,
)

# Attach SFT LoRA
model = PeftModel.from_pretrained(
    model,
    "vominhmanh/qwen3-4b-vihsd-lora",
)

model.eval()

---

🔁 Next Stage: DPO

This SFT LoRA adapter is intended to be used as the initial policy for:

• Direct Preference Optimization (DPO)
• Longer, higher-quality explanations
• Better alignment with human preferences

The DPO dataset is available at: 👉 https://huggingface.co/datasets/vominhmanh/vihsd-explainable-dpo

---

📜 License

Apache-2.0

---

🙏 Acknowledgements

• Original ViHSD dataset: htdung167/ViHSD
• Training framework: Unsloth
• Preference learning inspired by DPO (Rafailov et al.)