Datasets:

ExponentialScience
/

DLT-Sentiment-News

dataset_info:
  features:
    - name: timestamp
      dtype: string
    - name: title
      dtype: string
    - name: description
      dtype: string
    - name: text
      dtype: string
    - name: market_direction
      dtype:
        class_label:
          names:
            '0': neutral
            '1': bearish
            '2': bullish
    - name: engagement_quality
      dtype:
        class_label:
          names:
            '0': neutral
            '1': liked
            '2': disliked
    - name: content_characteristics
      dtype:
        class_label:
          names:
            '0': neutral
            '1': important
            '2': lol
    - name: vote_counts
      struct:
        - name: bearish
          dtype: int32
        - name: bullish
          dtype: int32
        - name: liked
          dtype: int32
        - name: disliked
          dtype: int32
        - name: important
          dtype: int32
        - name: lol
          dtype: int32
    - name: total_votes
      dtype: int32
    - name: source_url
      dtype: string
    - name: url
      dtype: string
    - name: total_tokens
      dtype: int64
  splits:
    - name: train
      num_bytes: 22639057
      num_examples: 23301
  download_size: 12118601
  dataset_size: 22639057
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
language:
  - en
tags:
  - DLT
  - Blockchain
  - Cryptocurrencies
  - Cryptocurrency
  - Bitcoin
  - Ethereum
  - XRP
  - Hedera
pretty_name: Distributed Ledger Technology (DLT) / Blockchain Sentiment News
size_categories:
  - 10K<n<100K

DLT-Sentiment-News

Dataset Description

Dataset Summary

DLT-Sentiment-News is a specialized sentiment analysis dataset for the Distributed Ledger Technology (DLT) domain. It addresses the lack of high-quality labeled data that captures domain-specific sentiment expressed by cryptocurrency community members.

The dataset contains 23,301 examples with 1.85 million tokens (average 79.51 tokens per example), spanning from January 2021 to May 2025. Each example includes cryptocurrency news headlines and descriptions with multi-dimensional sentiment labels crowdsourced from active community members on the CryptoPanic platform.

This dataset is part of the DLT-Corpus collection. For related datasets, see: https://huggingface.co/collections/ExponentialScience/dlt-corpus-68e44e40d4e7a3bd7a224402

Supported Tasks

Sentiment Analysis: Multi-dimensional sentiment classification for DLT and cryptocurrency content
Market Sentiment Studies: Analyzing how cryptocurrency communities perceive market-related news
Content Quality Assessment: Evaluating which content cryptocurrency users find valuable
Engagement Prediction: Understanding what drives positive or negative community engagement
Model Evaluation: Benchmarking domain-specific sentiment models

Languages

English (en)

Dataset Structure

Data Fields

Each example in the dataset contains the following fields:

title: Headline of the cryptocurrency news article
description: Brief description or summary of the article
text: Combined title and description text
timestamp: Date and time when the article was posted
market_direction: Sentiment about market direction (bullish, bearish, neutral)
engagement_quality: Community assessment of content importance (important, lol, neutral)
content_characteristics: User engagement type (liked, disliked, neutral)
vote_counts: Detailed breakdown of votes for each sentiment category
total_votes: Total number of community votes received
source_url: URL of the original news source
url: CryptoPanic URL for the article
total_tokens: Total number of tokens in the text

Label Distribution

The dataset includes three independent sentiment dimensions:

Market Direction:

bullish: Positive outlook on market/price movement
bearish: Negative outlook on market/price movement
neutral: Balanced or unclear market direction

Engagement Quality:

important: Content deemed significant by the community
lol: Content considered humorous or not serious
neutral: Standard content without strong quality signal

Content Characteristics:

liked: Positively received by the community
disliked: Negatively received by the community
neutral: Mixed or neutral community reception

Data Splits

This is a single corpus without predefined splits. Users should create their own train/validation/test splits based on their specific research needs. Consider temporal splits to avoid data leakage when studying market trends.

Dataset Creation

Curation Rationale

DLT-Sentiment-News was created to support sentiment analysis research in the DLT domain with data that reflects authentic community perspectives. Unlike general sentiment datasets, this captures:

Domain expertise: Labels from active cryptocurrency users with market knowledge
Multi-dimensional sentiment: Separate dimensions for market outlook, content quality, and engagement
Community consensus: Aggregated opinions from multiple users rather than single annotators
Market context: Sentiment tied to real cryptocurrency news and events

Source Data

Data Collection

The dataset was collected from CryptoPanic, a cryptocurrency news aggregation platform where community members vote on news articles across multiple sentiment categories.

Collection Details:

Data collected via CryptoPanic's free API between March and May 2025
Coverage period: January 2021 to May 2025
Only articles meeting minimum vote thresholds included (median minimum votes)
All content is publicly available news headlines and descriptions

Data Processing

The collection and labeling process involved:

Article retrieval: Collecting news articles with community votes from CryptoPanic
Vote normalization: Calculating vote percentages by total engagement for each article
Minimum threshold filtering: Excluding articles with insufficient community engagement (below median votes)
Percentile-based classification: Using 25th and 75th percentiles as boundaries to assign labels
Quality control: Ensuring balanced representation across sentiment categories

Annotations

Annotation Process

Crowdsourced Community Voting:

Active cryptocurrency community members on CryptoPanic vote on news articles
Users select from predefined sentiment categories for each dimension
Votes reflect genuine community sentiment and domain expertise

Label Assignment:

Percentile-based classification mitigates popularity bias
Articles below 25th percentile labeled as negative category
Articles above 75th percentile labeled as positive category
Articles between percentiles labeled as neutral category
Applied independently for each sentiment dimension

Who are the annotators?

Active cryptocurrency community members on the CryptoPanic platform. These annotators possess domain expertise and genuine interest in DLT/cryptocurrency news, providing more relevant sentiment labels than general crowdworkers.

Personal and Sensitive Information

This dataset contains only publicly available cryptocurrency news headlines and descriptions. No personal or confidential data is included. Individual voter information is not included - only aggregated vote counts and percentages are retained.

Considerations for Using the Data

Social Impact of Dataset

This dataset can enable:

Positive impacts: Better understanding of cryptocurrency community sentiment, improved market analysis tools, advancement of domain-specific NLP research, more accurate sentiment detection
Potential negative impacts: Could be misused for market manipulation, creating misleading investment systems, or amplifying market volatility through automated trading

Researchers should implement appropriate safeguards and ethical guidelines when working with this data.

Discussion of Biases

Potential biases include:

Platform bias: Only reflects CryptoPanic users, not the entire cryptocurrency community
Language bias: Only English-language news articles are included
Temporal bias: More recent years may have different sentiment patterns than earlier periods
User bias: Active voters may have different perspectives than passive readers
Source bias: Certain news sources may be over-represented
Market condition bias: Dataset may reflect specific market cycles (bull/bear markets)
Geographic bias: English-speaking regions and news sources are over-represented

Other Known Limitations

Temporal lag: Not suitable for real-time sentiment analysis
Market volatility: Sentiment may change rapidly after news publication
Vote manipulation: Despite filters, coordinated voting cannot be completely ruled out
Context dependency: Headlines lack full article context, which may affect sentiment interpretation
Evolving terminology: Cryptocurrency terminology and memes evolve rapidly
Static snapshot: Current version does not capture ongoing sentiment changes

Additional Information

Dataset Curators

Walter Hernandez Cruz, Peter Devine, Nikhil Vadgama, Paolo Tasca, Jiahua Xu

Licensing Information

CC-BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International)

This dataset is released under CC-BY-NC 4.0 for research purposes only.

Key terms:

Attribution required: You must give appropriate credit to the dataset creators
Non-commercial use: Commercial use is not permitted under this license
Academic research: The dataset is intended for academic and non-profit research

Legal basis:

Derived from publicly available CryptoPanic data with crowdsourced community annotations
Data collected via CryptoPanic's free API between March and May 2025
To the best of our knowledge, the Terms of Service at the time of collection (cryptopanic.com/terms/) contained no restrictions on academic research use or redistribution

For more information on CC-BY-NC 4.0, see: https://creativecommons.org/licenses/by-nc/4.0/

Acknowledgments

We thank the CryptoPanic platform and its community of users for making this dataset possible through their engagement and contributions to cryptocurrency news curation.

Citation Information

@article{hernandez2025dlt-corpus,
  title={DLT-Corpus: A Large-Scale Text Collection for the Distributed Ledger Technology Domain},
  author={Hernandez Cruz, Walter and Devine, Peter and Vadgama, Nikhil and Tasca, Paolo and Xu, Jiahua},
  year={2025}
}