Group : ForJustice3K

an AI model designed to detect hatespeech from message given. Based on WangchanBERTa. Trained with HateThaiSent, ThaiToxicityTweetCorpus.

**Group project for Artificial intelligence Class.

Usage

1. Environment setup

• Ensure Python 3.9+ is available.
• (Optional) Create and activate a virtual environment: python -m venv venv && source venv/bin/activate.
• Install dependencies: pip install -r requirements.txt.

If you're on Windows and want GPU acceleration, keep the optional torch-directml dependency; on other platforms it is safe to ignore if installation fails.

2. Prepare training data

• Place the labelled CSV in data/HateThaiSent.csv (default expected by the training script).
• The CSV must contain a Message column with text and a Hatespeech column with hatespeech/nonhatespeech labels.
• Optional: download and convert the Thai Toxicity Tweet corpus with python convert_thai_toxicity_tweet.py --output data/ThaiToxicityTweet_converted.csv; the training script consumes this file automatically (run training with --extra-data and no values to opt out).

3. Train the classifier

Fine-tune WangchanBERTa with the bundled training helper:

python train_model.py \
    --data data/HateThaiSent.csv \
    --output-dir models/wangchanberta-hatespeech \
    --epochs 2 \
    --batch-size 8

Key options:

• --device auto|cpu|cuda|mps|dml|directml: choose the execution device (DirectML unlocks Windows GPU acceleration).
• --freeze-encoder or --trainable-layer-count N: control how much of the backbone to fine-tune.
• --extra-data <path ...>: append additional CSVs that share the HateThaiSent schema; pass --extra-data with no values to disable the defaults.
• --max-length, --learning-rate, --gradient-accumulation: adjust training performance.
• --max-gpu-memory-fraction: leave VRAM headroom when working on limited GPUs such as Google Colab T4 instances.
• --lr-scheduler, --warmup-steps, --warmup-ratio: configure the learning-rate schedule and warmup strategy.
• --fp16 / --bf16: enable mixed-precision training on supported hardware.
• --gradient-checkpointing, --group-by-length: trade compute for lower memory and reduce padding overhead.
• --torch-compile: turn on PyTorch 2.x graph capture for potential throughput gains.

4. Run inference

After training finishes, score new text with age-aware post-processing:

python predict.py "ใส่ข้อความภาษาไทยที่นี่" --age 15 --model models/wangchanberta-hatespeech

The script prints a JSON object containing the raw probability, whether the content should be blocked for the provided age, and the policy thresholds that were applied.

5. Export metrics

Training writes evaluation metrics to models/wangchanberta-hatespeech/eval_metrics.json. Share this file or use it to monitor regression between fine-tuning runs.

6. Batch scoring helper

Use sample_predictions.py for quick qualitative checks or to score your own sentences in bulk:

python sample_predictions.py --model models/wangchanberta-hatespeech --input sentences.jsonl --output predictions.csv

The script prints per-sentence decisions, reports accuracy when expected labels are provided, and can export a CSV for downstream analysis. Default demo prompts now live in sample_sentences.jsonl; edit that file to tweak the built-in examples or pass a custom .jsonl/plain-text file through --input.

7. Next steps

• Integrate the exported model directory into the browser extension packaging workflow.
• Adjust age_policy.py if you need different moderation thresholds per age bracket.
• Run csv_output.py during data exploration to inspect label distribution and spot missing values.

Model credit

L. Lowphansirikul, C. Polpanumas, N. Jantrakulchai, and S. Nutanong, "WangchanBERTa: Pretraining transformer-based Thai language models," arXiv preprint arXiv:2101.09635, 2021. [Online]. Available: https://arxiv.org/abs/2101.09635.

Dataset credits

This project stands on the shoulders of public Thai-language moderation corpora. Please cite the sources below and respect the licences listed by the maintainers when redistributing or publishing results:

• HateThaiSent (data/HateThaiSent.csv). Data Science and Machine Learning Research Group (DSMLR), King Mongkut's University of Technology Thonburi. HateThaiSent: Hate speech and sentiment dataset for Thai social media (v1.0, 2021). Available at https://github.com/dsmlr/HateThaiSent. See the repository for detailed licensing terms.
• Thai Toxicity Tweet Corpus (data/ThaiToxicityTweet_converted.csv). National Electronics and Computer Technology Center (NECTEC) and Artificial Intelligence Association of Thailand (AIAT). Thai Toxicity Tweet Corpus (2020). Available via https://archive.org/download/ThaiToxicityTweetCorpus/data.zip with supporting documentation at https://github.com/tmu-nlp/ThaiToxicityTweetCorpus. Review the corpus licence before redistribution.