bert-cyberbullying-bahasa-classifier

A fine-tuned BERT multilingual classifier for detecting cyberbullying in Bahasa Indonesia. This model performs binary classification:

• 0 → non-bullying
• 1 → bullying

---

✅ Model Details

Property	Value
Model Type	BERT (base multilingual)
Task	Cyberbullying Detection (Text Classification)
Language	Bahasa Indonesia
Labels	0 — non-bullying, 1 — bullying
Framework	Hugging Face Transformers
Files	model.safetensors, config.json, tokenizer files

---

📚 Dataset

This model was trained using a combined dataset, consisting of:

• Indonesian cyberbullying dataset
• Additional toxic / abusive comment datasets
• Social media–style and chat–style text

Preprocessing steps:

• text normalization
• emoji removal
• punctuation cleanup
• lowercasing
• label encoding (0 / 1)

Dataset was balanced to reduce bias.

---

🧠 Training Information

• Base model: bert-base-multilingual-cased
• Epochs: 3–5
• Batch size: 16
• Optimizer: AdamW
• Learning rate: 2e-5
• Loss: Cross Entropy
• Train/Validation split: 80 / 20

Training was done on a 6GB GPU, optimized for low VRAM.

---

✅ How to Use

Python Example

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "zeltera/bert-cyberbullying-bahasa-classifier"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

text = "anjing lu jelek banget"
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits
    label = torch.argmax(logits, dim=1).item()

print("Prediction:", label)  # 1 = bullying

Example Predictions

Text	Output
"mampus lu biarin aja"	1 (bullying)
"kamu lagi dimana?"	0 (non-bullying)
"bodoh banget sih"	1 (bullying)
"nice job bro"	0 (non-bullying)

---

📈 Evaluation

Metric	Score
Accuracy	~0.90
F1 (macro)	~0.88
Precision	~0.89
Recall	~0.87

---

🗂️ Repository Contents

config.json
model.safetensors
tokenizer.json
tokenizer_config.json
special_tokens_map.json
vocab.txt
README.md

---

🔧 Intended Use

• AI chatbots (moderation / filtering)
• Social media comment analysis
• Cyberbullying detection systems
• Student safety applications
• Research on toxicity detection

---

⚠️ Limitations

• Limited sarcasm detection
• May misclassify unseen slang
• Works best on Indonesian text
• Not suitable for legal or high-risk decisions

---

📜 License

MIT License

---

👤 Author

Model trained and published by @zeltera Built using Hugging Face Transformers + PyTorch. Contact instagram @gnwnadiwjy