Bluesky posts

Approximately 9 million public Bluesky posts, processed and cleaned for machine learning research and experimentation. The dataset has been normalized and filtered to remove duplicates, with sensitive information replaced by placeholders.

This dataset isn't directly from Bluesky itself. It's a processed version of the Roronotalt/bluesky-ten-million dataset.

Dataset Details

• Size: Approximately 9 million posts
• Format: JSON Lines (.jsonl)
• Source: Roronotalt/bluesky-ten-million

Preprocessing

The dataset has undergone the following preprocessing steps:

1. Normalization of bullet points, hyphenated words, quotation marks, unicode, and whitespace
2. Replacement of emails, phone numbers, URLs, and user handles with placeholders
3. Removal of duplicate posts
4. Language detection using fasttext, gcld3, and Lingua (if any classifier gives >95% confidence)

File Structure

The dataset is organized into separate files based on detected languages:

• all.jsonl: Contains posts in all languages
• [language_code].jsonl: Individual files for each detected language
• unknown.jsonl: Posts with undetected or low-confidence language

Each file is compressed using zstd.

Data Format

Each line in the JSON Lines files represents a single post with the following structure:

{"text": "hello world :D", "language": "en"}

• text: The processed text content of the post
• language: The detected language code (ISO 639-1), the language code with its script, or "unknown"

Ethical Considerations

This dataset contains public posts from Bluesky Social. While personal identifiers have been replaced with placeholders, users should be aware of potential biases and privacy concerns when using this data for research or applications.

Limitations

• Very bad language detection for short or ambiguous posts especially when the post includes emojis, names, hashtags, mispellings, and O-D-D-L-Y F-O-R-M-A-T-T-E-D TEXT~!!!
• The dataset may not be representative of the entire Bluesky user base or social media in general
• Content moderation status is not included, so the dataset may contain inappropriate or offensive material
• The detailed information on each post and their authors have been removed in order to concentrate on the post content itself