README.md

metadata

language:
  - tr
tags:
  - text-classification
  - bert
  - offensive-language-detection
  - turkish
  - boun-tabilab
datasets:
  - offenseval-tr
metrics:
  - accuracy
  - f1
model-index:
  - name: atahanuz/bert-offensive-classifier
    results:
      - task:
          type: text-classification
          name: Text Classification
        dataset:
          name: OffensEval-2020-TR
          type: offenseval-tr
        metrics:
          - name: Accuracy
            type: accuracy
            value: 0.936
          - name: F1
            type: f1
            value: 0.912
base_model: boun-tabilab/TabiBERT

Turkish Offensive Language Classifier (BERT)

This model is a fine-tuned version of boun-tabilab/TabiBERT trained on the OffensEval-2020-TR dataset. It is designed to perform binary classification to detect offensive language in Turkish text.

📊 Model Details

Feature	Description
Model Architecture	BERT (Base Uncased Turkish - TabiBERT)
Task	Binary Text Classification (Offensive vs. Not Offensive)
Language	Turkish (tr)
Dataset	OffensEval 2020 (Turkish Subtask)
Trained By	atahanuz

🚀 Usage

The easiest way to use this model is via the Hugging Face pipeline.

Method 1: Using the Pipeline (Recommended)

from transformers import pipeline

classifier = pipeline("text-classification", model="atahanuz/bert-offensive-classifier")

text = "Bu harika bir filmdi, çok beğendim."
result = classifier(text)[0]

label = "Offensive" if result['label'] == "LABEL_1" else "Not Offensive"

print(f"Prediction: {label} (Score: {result['score']:.4f})")

# Prediction: Not Offensive (Score: 1.0000)

Method 2: Manual PyTorch Implementation

If you need more control over the tokens or logits, use the standard AutoModel approach:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# 1. Load model and tokenizer
model_name = "atahanuz/bert-offensive-classifier"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# 2. Define label mapping
id2label = {0: "NOT", 1: "OFF"}

# 3. Tokenize and predict
text = "Bu harika bir filmdi, çok beğendim." # Example text
inputs = tokenizer(text, return_tensors="pt", truncation=True)

with torch.no_grad():
    logits = model(**inputs).logits

# 4. Get results
predicted_class_id = logits.argmax().item()
predicted_label = id2label[predicted_class_id]
confidence = torch.softmax(logits, dim=1)[0][predicted_class_id].item()

print(f"Text: {text}")
print(f"Prediction: {predicted_label} (Confidence: {confidence:.4f})")

🏷️ Label Mapping

The model outputs the following labels:

Label ID	Label Name	Description
0	NOT	Not Offensive - Normal, non-hateful speech.
1	OFF	Offensive - Contains insults, threats, or inappropriate language.

📝 Example Predictions

Text	Label	Prediction
"Bu filmi çok beğendim, oyunculuklar harikaydı."	NOT	Non-Offensive
"Beynini kullanmayı denesen belki anlarsın."	OFF	Offensive
"Maalesef bu konuda sana katılamıyorum."	NOT	Non-Offensive
"Senin gibi aptal insanlar yüzünden bu haldeyiz."	OFF	Offensive

📈 Performance

The model was evaluated on the test split of the OffensEval-2020-TR dataset (approx. 3,500 samples).

• Accuracy: 93.6%
• F1 Score: 91.2%

Dataset Statistics

• Training Samples: 31,277
• Test Samples: 3,529

⚠️ Limitations and Bias

• Context Sensitivity: Like many BERT models, this classifier may struggle with sarcasm or offensive language that depends heavily on context not present in the input sentence.
• Dataset Bias: The model is trained on social media data (OffensEval). It may reflect biases present in that specific dataset or struggle with formal/archaic Turkish.
• False Positives: Certain colloquialisms or "tough love" expressions might be misclassified as offensive.

📚 Citation

If you use this model, please cite the TabiLAB model and OffensEval paper:

@misc{Türker2025Tabibert,
      title={TabiBERT: A Large-Scale ModernBERT Foundation Model and Unified Benchmarking Framework for Turkish}, 
      author={Melikşah Türker and Asude Ebrar Kızıloğlu and Onur Güngör and Susan Üsküdarlı},
      year={2025},
      eprint={2512.23065},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2512.23065}, 
}

@inproceedings{zampieri-etal-2020-semeval,
    title = "{SemEval}-2020 Task 12: Multilingual Offensive Language Identification in Social Media ({OffensEval} 2020)",
    author = "Zampieri, Marcos  and
      Nakov, Preslav  and
      Rosenthal, Sara  and
      Atanasova, Pepa  and
      Karadzhov, Georgi  and
      Mubarak, Hamdy  and
      Derczynski, Leon  and
      Pym, Z",
    booktitle = "Proceedings of the Fourteenth Workshop on Semantic Evaluation",
    year = "2020",
    publisher = "International Committee for Computational Linguistics",
}