vominhmanh/qwen3-4b-vihsd

qwen3-4b-instruct-2507-vihsd-explainable

Overview

This repository provides a LoRA-finetuned and merged version of
unsloth/Qwen3-4B-Instruct-2507 for Vietnamese toxic content moderation (ViHSD).

Unlike standard classification models, this model is trained to generate structured, explainable outputs in JSON format, including:

• label: CLEAN | OFFENSIVE | HATE
• explanation: short Vietnamese rationale
• evidence: verbatim substrings supporting the decision

The primary goal of this work is task-aligned, explainable moderation, not only raw label prediction.

---

Base model

• Architecture: Qwen3-4B-Instruct-2507
• Provider: unsloth
• Type: Instruction-tuned causal language model

---

Fine-tuned model

• Fine-tuning method: LoRA (parameter-efficient fine-tuning)
• Trainable parameters: ~0.8% of total parameters
• Output format: Strict JSON (label + explanation + evidence)
• Language: Vietnamese

---

Dataset

• Dataset name: vominhmanh/vihsd-explainable
• Derived from: ViHSD (Vietnamese Hate Speech Detection)
• Splits:
  • Train: 24,048 samples
  • Validation: 2,672 samples
  • Test: 6,680 samples
• Annotations:
  • label (CLEAN / OFFENSIVE / HATE)
  • explanation (Vietnamese rationale)
  • evidence (verbatim toxic spans, empty for CLEAN)

---

Training setup

• Objective: Causal language modeling (NLL loss)
• Prompt format: Chat-style (user → assistant JSON response)
• LoRA configuration:
  • r = 16
  • alpha = 32
  • dropout = 0
• Precision: bf16
• Optimizer: AdamW (8-bit)
• Epochs: 2
• Sequence length: 512
• Framework: Unsloth + TRL SFTTrainer

---

Evaluation protocol

Metric

• Macro F1 (3-class) over labels {CLEAN, OFFENSIVE, HATE}

Evaluation procedure

• Base model and fine-tuned model were evaluated using:
  • identical prompt
  • identical JSON parsing logic
  • identical evaluation script
• Evaluation performed on 6,680 test samples

---

Results

Model	Macro F1 (3-class)
Base: Qwen3-4B-Instruct-2507	0.5085
Fine-tuned (this work)	0.6370

Interpretation

• Fine-tuning yields a +12.85 absolute Macro F1 improvement over the base model.
• This demonstrates that the base instruction-tuned model is not well-aligned with ViHSD-specific moderation definitions.
• LoRA fine-tuning significantly improves task alignment and label discrimination, even with <1% trainable parameters.

---

Why this improvement matters

The improvement is not only quantitative but also qualitative:

1. Better label discrimination
   • Reduced confusion between OFFENSIVE and HATE
2. Stronger alignment with ViHSD definitions
   • Fine-tuned model follows dataset-specific guidelines
3. Explainable outputs
   • Consistent explanations and evidence spans
4. Structured generation
   • Reliable JSON output suitable for downstream pipelines

---

Limitations

• Evaluation focuses on label-level Macro F1; explanation and evidence quality are not fully captured by this metric.
• Explanations are generated heuristically and may still contain errors.
• Not intended for fully automated moderation without human review.

---

Intended use

• Research on Vietnamese toxic content moderation
• Explainable AI for content review systems
• Human-in-the-loop moderation pipelines

---

License

Please verify compatibility with:

• Base model license: unsloth/Qwen3-4B-Instruct-2507
• Dataset license: vominhmanh/vihsd-explainable

---

Citation

If you use this model, please cite:

• ViHSD dataset
• Qwen3-4B-Instruct
• This repository