What dataset should I use for sentiment analysis with labeled text?

A good dataset depends on your goal, language, and whether you want binary, 3-class, or fine-grained sentiment labels. Here are the most commonly used labeled text sentiment datasets:

Good general-purpose choices

• IMDb Large Movie Reviews

  • Labels: positive / negative
  • Size: 50,000 reviews
  • Best for: binary sentiment classification with longer texts
  • Why use it: very standard and easy to work with

• SST-2 (Stanford Sentiment Treebank)

  • Labels: positive / negative
  • Size: smaller than IMDb
  • Best for: sentence-level sentiment
  • Why use it: commonly used in NLP benchmarks and works well for short text

For multi-class sentiment

• SST-5

  • Labels: very negative, negative, neutral, positive, very positive
  • Best for: fine-grained sentiment classification

• Amazon Reviews

  • Labels: often 1–5 stars
  • Best for: large-scale training and domain-specific sentiment
  • Why use it: lots of data and many product domains

For social media / short informal text

• Sentiment140

  • Labels: positive / negative
  • Text: tweets
  • Best for: social media sentiment

• TweetEval

  • Labels: depends on task split, often positive / negative / neutral
  • Best for: Twitter/X-style short text

For multilingual sentiment

• Multilingual Amazon Reviews
• XED / multilingual sentiment datasets
• SemEval multilingual tasks
  Good if you need non-English sentiment analysis.

My recommendation

• If you want a simple, standard starting point: IMDb or SST-2
• If you want fine-grained sentiment: SST-5
• If you want tweets/social media: Sentiment140 or TweetEval
• If you want lots of data: Amazon Reviews

If you want, I can also recommend the best dataset for your specific use case (e.g., product reviews, tweets, customer support, multilingual, or transformer fine-tuning).