---
license: mit
datasets:
- winvoker/turkish-sentiment-analysis-dataset
language:
- tr
base_model:
- answerdotai/ModernBERT-large
---

Here's an updated **Model Card** in a **README format** based on the training results and the model you've used (ModernBERT-large for Turkish sentiment analysis):

```markdown
# Turkish Sentiment ModernBERT-large
```
This is a fine-tuned **ModernBERT-large** model for **Turkish Sentiment Analysis**. The model was trained on the `winvoker/turkish-sentiment-analysis-dataset` and is designed to classify Turkish text into sentiment categories such as positive, negative, and neutral.

## Model Overview

- **Model Type**: ModernBERT (BERT variant)
- **Task**: Sentiment Analysis
- **Languages**: Turkish
- **Dataset**: [winvoker/turkish-sentiment-analysis-dataset](https://huggingface.co/datasets/winvoker/turkish-sentiment-analysis-dataset)
- **Labels**: Positive, Negative, Neutral
- **Fine-Tuning**: Fine-tuned for sentiment classification.

## Performance Metrics

The model was trained for **4 epochs** with the following results:

| Epoch | Training Loss | Validation Loss | Accuracy | F1 Score |
|-------|---------------|-----------------|----------|----------|
| 1     | 0.2884        | 0.1133          | 95.72%   | 92.18%   |
| 2     | 0.1759        | 0.1050          | 96.24%   | 93.33%   |
| 3     | 0.0633        | 0.1233          | 96.14%   | 93.19%   |
| 4     | 0.0623        | 0.1213          | 96.14%   | 93.19%   |

- **Training Loss**: Measures how well the model fits the training data.
- **Validation Loss**: Measures how well the model generalizes to unseen data.
- **Accuracy**: Percentage of correct predictions over all examples.
- **F1 Score**: A balanced metric between precision and recall, accounting for both false positives and false negatives.

## Model Inference Example

You can use this model for sentiment analysis of Turkish text. Here’s an example of how to use it:

```python
from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

# Load the pre-trained model and tokenizer
model_name = "bayrameker/Turkish-sentiment-ModernBERT-large"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Example texts for prediction
texts = ["bu ürün çok iyi", "bu ürün berbat"]

# Tokenize the inputs
inputs = tokenizer(texts, padding=True, truncation=True, return_tensors="pt")

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

# Get the predicted sentiment labels
predictions = torch.argmax(logits, dim=-1)
labels = ["Negative", "Neutral", "Positive"]  # Adjust based on your label mapping
for text, pred in zip(texts, predictions):
    print(f"Text: {text} -> Sentiment: {labels[pred.item()]}")
```

### Example Output:

```
Text: bu ürün çok iyi -> Sentiment: Positive
Text: bu ürün berbat -> Sentiment: Negative
```

## Installation

To use this model, install the following dependencies:

```bash
pip install transformers
pip install torch
pip install datasets
```

## Model Card

- **Model Name**: Turkish-sentiment-ModernBERT-large
- **Hugging Face Repo**: [Link to Model Repository](https://huggingface.co/bayrameker/Turkish-sentiment-ModernBERT-large)
- **License**: MIT (or any applicable license you choose)
- **Author**: Bayram Eker
- **Date**: 2024-12-21

## Training Details

- **Model**: ModernBERT-large
- **Framework**: PyTorch
- **Training Time**: Approximately 50 minutes (4 epochs)
- **Batch Size**: 64
- **Learning Rate**: 8e-5
- **Optimizer**: AdamW
- **Mixed Precision**: bf16 for A100 GPU

## Acknowledgments

- The model was trained on the `winvoker/turkish-sentiment-analysis-dataset` dataset.
- Special thanks to the Hugging Face community and the contributors to the transformers library.
- Thanks to all contributors of the dataset and pretrained models.

## Future Work

- Expand the model with more complex sentiment labels (e.g., multi-class sentiments, aspect-based sentiment analysis).
- Fine-tune the model on a larger, more diverse dataset for better generalization across various domains.

## License

This model is licensed under the MIT License. See the LICENSE file for more details.