Arshia82sbn/youtube-sentiment-dataset

YouTube Comments Sentiment Dataset

A comprehensive, large-scale dataset featuring 1,032,225 English YouTube comments, curated and labeled for 3-class sentiment analysis (Negative, Neutral, and Positive). This dataset is optimized for training, evaluating, and fine-tuning Transformer-based NLP models and sentence encoders.

🔗 Related Resources

• —Hugging Face Dataset: Arshia82sbn/youtube-sentiment-dataset
• —Hugging Face Model: Arshia82sbn/youtube-sentiment-classifier-english-mpnet

Dataset Description

Class Distribution

Split

Columns Definition

• —VideoTitle — Original title of the source YouTube video.
• —CommentText — Raw user comment text.
• —Sentiment — Categorical text label (Negative, Neutral, Positive).
• —Sentiment_label — Numeric target label (0 for Negative, 1 for Neutral, 2 for Positive).
• —clean_comment_text — Normalized and noise-reduced comment text.
• —tokens_comment_text — Lemmatized/tokenized list representation of the comment.
• —clean_title — Normalized video title text.
• —token_title — Lemmatized/tokenized list representation of the video title.

Usage

Using Hugging Face datasets

from datasets import load_dataset

# Load directly from the Hugging Face Hub
dataset = load_dataset("Arshia82sbn/youtube-sentiment-dataset")

print(dataset)
print("Sample record:", dataset["train"][0])

Using pandas

import pandas as pd

# If loaded from local CSV or Parquet export
df = pd.read_csv("youtube-comments-sentiment.csv")
print(f"Total samples: {len(df)}")
print(df.head())

Model Inference

A pre-trained MPNet classifier built on this dataset is available:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Arshia82sbn/youtube-sentiment-classifier-english-mpnet")
embeddings = model.encode(["This video was incredibly helpful!", "Terrible content, waste of time."])
print(embeddings.shape)

License

This dataset is released for academic, research, and educational purposes.