metadata

license: ecl-2.0
task_categories:
  - text-classification
language:
  - en
tags:
  - public
  - text
  - education
  - sentiment-analysis
pretty_name: TweetFeels 1m6
size_categories:
  - 1M<n<10M
configs:
  - config_name: dataset
    data_files:
      - split: train
        path:
          - tweets_1m6_processed.csv

TweetFeels 1m6

Collection of over a million English-language tweets harvested through the Twitter API in 2009. Each tweet carries a sentiment label automatically inferred from the presence of positive or negative emoticons: 0 for negative and 4 for positive (the original release also encodes 2 for neutral, though this class is sparsely populated).

The dataset is distributed as a single CSV file with six tab-separated fields:

target – sentiment polarity (0, 2, 4)
ids – unique tweet ID
date – UTC timestamp of posting
flag – search query that matched the tweet, or “NO_QUERY”
user – screen name of the posting account
text – full tweet text, emoticons and URLs preserved

No manual annotation was performed; labels were assigned via distant supervision based on emoticon presence. The corpus covers diverse topics and informal language patterns typical of Twitter, making it a standard benchmark for large-scale sentiment analysis and social-media text-mining tasks.

Acknowledgements: The dataset is available through Kaggle—Sentiment140.