📊 YıldızSezar: Turkish E-Commerce Reviews (Real + LLaMA Synthetic)

This dataset is a comprehensive collection of Turkish customer reviews from e-commerce platforms, labeled with 1-to-5 star ratings. It was developed to train highly accurate multi-class sentiment analysis models.

It is the official dataset for the peer-reviewed paper: A Star Rating-Based Approach in BERT-Based Sentiment Analysis of Customer Feedback.

📌 Dataset Overview

One of the biggest challenges in analyzing e-commerce reviews is the class imbalance—users mostly leave either 5-star (very happy) or 1-star (very angry) reviews.

To solve this, I augmented the real-world dataset by generating over 900,000 synthetic Turkish reviews specifically targeting the minority classes (2, 3, and 4 stars) using LLaMA-8B-DPO. The resulting dataset is highly balanced and morphologically complex.

• Total Synthetic Samples: ~900,000
• Task: 5-Class Sentiment Analysis / Star Rating Prediction
• Language: Turkish

📂 Data Splits

The dataset is ready-to-use and pre-split for training, validation, and testing.

💻 Quick Usage

from datasets import load_dataset

dataset = load_dataset("ilkayO/yildizsezar-turkish-reviews")
print(dataset['train'][0])
# Example Output: {'review_text': 'ürün çok kaliteli tavsiye ederim', 'star_rating': 5}

⚖️ License

This dataset is published under the CC BY-NC 4.0 license for research and academic purposes.

(If you are interested in commercial applications or custom NLP data pipelines for your enterprise, please check the contact details on my Model Page.)