engin1123/keloglan-turkish-sentiment-analysis-dataset

🏰 Keloğlan Turkish Sentiment Analysis Dataset

This dataset represents one of the largest and most comprehensive sentiment analysis collections for the Turkish language, containing over 630,000 unique samples.

It was meticulously constructed by merging, cleaning, and deduplicating several major open-source Turkish sentiment datasets to train the Keloğlan model series.

📊 Dataset Statistics

• —Total Unique Samples: 631,166
• —Training Set: 599,607
• —Validation Set: 31,559
• —Label Distribution:
• —0 (Negative): Mixed sources (Reviews + Irony)
• —1 (Neutral): Objective statements, mixed feelings
• —2 (Positive): High satisfaction reviews

🧩 Data Sources & Construction

This dataset is a unification of the following sources, processed to create a standardized format:

1. 1.WhiteAngelss/Turkce-Duygu-Analizi-Dataset: E-commerce reviews (mapped 1-5 stars to Neg/Neu/Pos).
2. 2.Winvoker/turkish-sentiment-analysis-dataset: A large collection of movie and product reviews.
3. 3.Maydogan/Turkish_SentimentAnalysis_TRSAv1: A well-known benchmark dataset.
4. 4.Turkish Product Reviews: Additional diverse product feedback.

🛠️ Preprocessing Pipeline

To ensure high quality and prevent data leakage:

1. 1.Standardization: All labels were mapped to a unified {0: Negative, 1: Neutral, 2: Positive} schema.
2. 2.Deduplication: Aggressive deduplication was applied. From an initial pool of 1.26 million rows, we removed exact text duplicates, resulting in 631k unique samples. This prevents the model from memorizing repetitive comments (e.g., "Teşekkürler", "Kargo hızlıydı").
3. 3.Cleaning: Removed extremely short texts (len < 3) and noise.
4. 4.Shuffling: The dataset is fully shuffled to ensure balanced batches during training.

🚀 How to Use

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("engin1123/keloglan-turkish-sentiment-analysis-dataset")

# Access splits
train_data = dataset['train']
val_data = dataset['validation']

print(train_data[0])
# Output: {'text': 'Ürün gayet başarılı...', 'label': 2}

⚖️ License

This dataset is a compilation of open-source datasets. Users should refer to the original licenses of the source datasets (Winvoker, WhiteAngelss, TRSAv1). The compilation itself is shared under CC-BY-NC 4.0.

👨‍💻 Maintainer

Maintained by Engin Yazilan for the Keloğlan project.