ragibhasan/coke_studio_bangla_sentiment

Coke Studio Bangla YouTube Comments Sentiment Dataset

A rich collection of 57,592 YouTube comments across 33 songs and 4 seasons of Coke Studio Bangla, labeled with sentiment classes (positive, neutral, negative), model confidence scores, language scripts (Bengali, Latin/Banglish, mixed), topics, and engagement metrics.

---

Dataset Summary

• Total Comments: 57,592
• Unique Songs: 33 tracks across Coke Studio Bangla Seasons 1 to 4
• Language / Scripts:
  • Latin / Banglish / English: 32,098 (55.7%)
  • Bengali script (বাংলা): 23,877 (41.5%)
  • Mixed script: 1,518 (2.6%)
  • Emoji or other: 99 (0.2%)
• Sentiment Distribution:
  • Positive: 36,167 (62.8%)
  • Neutral: 10,945 (19.0%)
  • Negative: 10,480 (18.2%)
• Sentiment Model: cardiffnlp/twitter-xlm-roberta-base-sentiment
• Configurations:
  • default: Complete 26-column schema with raw text, clean text, metadata, topic classification, and sentiment scores.
  • curated: Streamlined 16-column schema focused on clean text and sentiment classification.

---

Quickstart & Usage

Install the datasets library:

pip install datasets pyarrow

Loading the Default (Full) Dataset

from datasets import load_dataset

# Load default configuration (26 features)
ds = load_dataset("ragibhasan/coke_studio_bangla_sentiment")

print(ds["train"][0])

Loading the Curated Dataset

# Load curated configuration (16 features)
ds_curated = load_dataset("ragibhasan/coke_studio_bangla_sentiment", "curated")

print(ds_curated["train"][0])

---

Dataset Structure

Configurations

Configuration	Features	Description
default	26	Full metadata: comment_id, video_id, song, season, topic, subtopic, text_clean, text_raw, emojis, script, n_chars, n_words, n_emoji, votes, is_reply, parent_cid, has_heart, published_at, author_id, p_negative, p_neutral, p_positive, sentiment, confidence, sentiment_model, needs_review
curated	16	Focused on modeling: comment_id, video_id, song, season, text_clean, script, emojis, votes, is_reply, published_at, sentiment, confidence, p_negative, p_neutral, p_positive, needs_review

Field Descriptions

• comment_id (string): Unique identifier for the YouTube comment.
• video_id (string): YouTube video ID.
• song (string): Name of the Coke Studio Bangla song (e.g., Bulbuli, Kotha Koiyo Na, Murir Tin, Long Distance Love).
• season (int64): Coke Studio Bangla season number (1, 2, 3, 4).
• topic (string): Broad comment theme/topic (e.g., music, artist, production).
• subtopic (string): Granular subtopic classification.
• text_clean (string): Preprocessed comment text (normalized, stripped of URLs/excess whitespace).
• text_raw (string): Original unfiltered comment text as retrieved from YouTube.
• emojis (string): Emojis extracted from the comment text.
• script (string): Detected writing script (bangla, latin, mixed, emoji_or_other).
• n_chars (int64): Character length of the comment.
• n_words (int64): Word count of the comment.
• n_emoji (int64): Number of emoji characters present.
• votes (int64): YouTube like count on the comment.
• is_reply (bool): Whether the comment is a reply (True) or a top-level comment (False).
• parent_cid (string): Parent comment ID (if this is a reply).
• has_heart (bool): Whether the comment was hearted by the creator channel.
• published_at (timestamp): UTC publication timestamp.
• author_id (string): Anonymized hash of the comment author ID.
• p_negative (float64): Probability of negative sentiment predicted by XLM-RoBERTa.
• p_neutral (float64): Probability of neutral sentiment predicted by XLM-RoBERTa.
• p_positive (float64): Probability of positive sentiment predicted by XLM-RoBERTa.
• sentiment (string): Predicted sentiment label (positive, neutral, negative).
• confidence (float64): Model prediction confidence (max probability).
• sentiment_model (string): Sentiment analysis model checkpoint used.
• needs_review (bool): Flag indicating low confidence or ambiguous predictions for human review.

---

Dataset Breakdown & Insights

Sentiment Distribution

• Positive: 36,167 (~62.8%)
• Neutral: 10,945 (~19.0%)
• Negative: 10,480 (~18.2%)

Season Breakdown

Season	Comments
Season 2	20,561
Season 1	15,537
Season 3	14,553
Season 4	6,941

Top Songs Represented

1. Bulbuli (2,598 comments)
2. Long Distance Love (2,551 comments)
3. Kotha Koiyo Na (2,423 comments)
4. Murir Tin (2,403 comments)
5. Moha Jadu (2,235 comments)
6. ...and 28 other popular tracks.

Visualizations & Insights

Sentiment Distribution by Song

Net Sentiment Trend

Net Sentiment by Season

---

Use Cases

1. Code-Mixed & Multilingual Sentiment Analysis: Benchmark models on Bengali text written in Bengali script vs. Banglish (Bengali in Latin script).
2. Music & Entertainment Domain NLP: Analyze reception, fan sentiment, and audience engagement across music releases.
3. Low-Resource & Regional NLP: Fine-tune multilingual transformers (XLM-RoBERTa, BanglaBERT, IndicBERT, mT5) for social media text processing.
4. Topic & Engagement Modeling: Study the relationship between comment likes, reply depth, emojis, and overall sentiment.

---

Citation & Attribution

If you use this dataset in your research or project, please cite:

@dataset{coke_studio_bangla_sentiment,
  author       = {Hasan, Ragib and Sen Toma, Indrani and Galib, Asadulla Hel},
  title        = {Coke Studio Bangla YouTube Comments Sentiment Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  version      = {226b36f},
  doi          = {10.57967/hf/10724},
  url          = {https://huggingface.co/datasets/ragibhasan/coke_studio_bangla_sentiment}
}

License

This dataset is distributed under the Apache License 2.0. YouTube comments remain subject to YouTube's Terms of Service.