nanelimon/bert-base-turkish-offensive

About the model

This model is designed for text classification, specifically for identifying offensive content in Turkish text. The model classifies text into five categories: INSULT, OTHER, PROFANITY, RACIST, and SEXIST.

Model Metrics

	INSULT	OTHER	PROFANITY	RACIST	SEXIST
Precision	0.901	0.924	0.978	1.000	0.980
Recall	0.920	0.980	0.900	0.980	1.000
F1 Score	0.910	0.9514	0.937	0.989	0.990

• F-Score: 0.9559690799177005
• Recall: 0.9559999999999998
• Precision: 0.9570284225256961
• Accuracy: 0.956

Training Information

• Device : macOS 14.5 23F79 arm64 | GPU: Apple M2 Max | Memory: 5840MiB / 32768MiB
• Training completed in 0:22:54 (hh:mm:ss)
• Optimizer: AdamW
• learning_rate: 2e-5
• eps: 1e-8
• epochs: 10
• Batch size: 64

Dependency

pip install torch torchvision torchaudio
pip install tf-keras  
pip install transformers  
pip install tensorflow

Example

from transformers import AutoTokenizer, TFAutoModelForSequenceClassification, TextClassificationPipeline

# Load the tokenizer and model
model_name = "nanelimon/bert-base-turkish-offensive"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = TFAutoModelForSequenceClassification.from_pretrained(model_name)

# Create the pipeline
pipe = TextClassificationPipeline(model=model, tokenizer=tokenizer, return_all_scores=True, top_k=2)

# Test the pipeline
print(pipe('Bu bir denemedir hadi sende dene!'))

Result;

[[{'label': 'OTHER', 'score': 1.000}, {'label': 'INSULT', 'score': 0.000}]]

• label= It shows which class the sent Turkish text belongs to according to the model.
• score= It shows the compliance rate of the Turkish text sent to the label found.

Authors

• Seyma SARIGIL: seymasargil@gmail.com

License

gpl-3.0

Free Software, Hell Yeah!