bert-turkish-sentiment-analysis

Model Description

This model is a fine-tuned version of dbmdz/bert-base-turkish-cased on the winvoker/turkish-sentiment-analysis-dataset.

• Shuffle function was applied during preprocessing.
• Training dataset: 229,483 samples
• Test dataset: 57,371 samples
• Total dataset: 286,854 samples

This model is designed for Turkish sentiment analysis, classifying texts as either Positive (1) or Negative (0).

Training Details:

Training dataset

• Dataset: winvoker/turkish-sentiment-analysis-dataset
• Training: 229,483 samples
• Testing: 57,371 samples
• Preprocessing: shuffle and truncate/pad to 128 tokens

Training Procedure

Fine-tuned using Hugging Face Transformers Trainer

• Learning rate: 2e-5
• Batch size: 16
• Epochs: 2
• Optimizer: AdamW
• Weight decay: 0.01

Evaluation

• Eval Loss: 0.1505
• Accuracy: 0.9614
• F1 Score: 0.9767
• Epoch: 2

Label Mapping

• LABEL_1 = Positive (1)

• LABEL_0 = Negative (0)

• Developed by: Yusuf Altunbaş

• Model type: Sequence Classification

• Language(s): Turkish

• License: MIT

• Finetuned from model: dbmdz/bert-base-turkish-cased

Model Sources

• Repository: GitHub Notebook

Uses

Direct Use

• The model can be directly used for predicting the sentiment of Turkish text using the Hugging Face transformers library.

Downstream Use

• Can be integrated into larger NLP pipelines, sentiment analysis dashboards, or recommendation systems.

Out-of-Scope Use

• This model is not trained for neutral or multi-class sentiment beyond Positive/Negative.
• Not suitable for languages other than Turkish.

Bias, Risks, and Limitations

• The model is trained on winvoker/turkish-sentiment-analysis-dataset.
• It may reflect biases present in the dataset.
• Misuse or misinterpretation of predictions can occur; always review results in context.

How to Get Started with the Model

Using Hugging Face Pipeline

from transformers import pipeline

text = "Bugün çok mutluyum!"
model_id = "yusufalt46/bert-turkish-sentiment-analysis"

classifier = pipeline("text-classification", model=model_id, tokenizer=model_id)
preds = classifier(text)
print(preds)

# Example output:
# [{'label': 'LABEL_1', 'score': 0.9823}]  # 1 = Positive

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("yusufalt46/bert-turkish-sentiment-analysis")
model = AutoModelForSequenceClassification.from_pretrained("yusufalt46/bert-turkish-sentiment-analysis")

text = "Bu ürün harika!"
inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=128)
outputs = model(**inputs)
pred = torch.argmax(outputs.logits, dim=-1)
print(pred.item())  # 1: Positive, 0: Negative

Yusuf Altunbaş, bert-turkish-sentiment-analysis, Hugging Face Model Hub, 2025