lukeslp/bluesky-alt-text

Bluesky Alt Text: Frozen April 2026 Snapshot

Frozen historical snapshot. These files were collected in April 2026 and will not be extended into a longitudinal series. The 279K-row corpus was deliberately selected from accounts with high alt-text adoption, so it must not be interpreted as a representative platform adoption estimate. The 14.5-hour Jetstream file is a short observed window, not a durable census.

Rows contain author handles and DIDs, post URIs and CIDs, post text, image references/URLs, collection cursors, and complete raw AT Protocol records. They are identifiable public-post records, not anonymous data. Bluesky users retain rights in their text; this repository does not claim authority to relicense their descriptions under CC BY. See the current, minimized Bluesky Accessibility Observatory for aggregate trends and a deletion-responsive pseudonymous sample.

279,196 image alt-text pairs from 489 deliberately selected Bluesky accounts, collected via public AT Protocol APIs. One row represents one image, not one post.

Includes a 125K-row firehose sample (14.5 hours of live network traffic) for contrastive analysis.

Overview

• Collection period: April 2026
• Total rows: 279,196 (corpus) + 125,645 (firehose) = 404,841
• Unique authors: 489 (corpus), 34,785 (firehose)
• Mean alt text length: 203 characters
• Median: 127 characters
• 60.5% of entries exceed 100 characters
• Languages: English (94%), German, French, Portuguese, Spanish, Japanese, 20+ others
• Collection method: AT Protocol getAuthorFeed (corpus), Jetstream WebSocket (firehose)

Download

GitHub Release (recommended for bulk download):

• corpus.jsonl.gz (89 MB)
• corpus.csv.gz (77 MB)
• firehose.jsonl.gz (49 MB)

Quick Start

import pandas as pd

corpus = pd.read_json("corpus.jsonl.gz", lines=True)
firehose = pd.read_json("firehose.jsonl.gz", lines=True)

print(f"Corpus: {len(corpus):,} rows, mean alt {corpus.alt_text.str.len().mean():.0f} chars")
print(f"Firehose: {len(firehose):,} rows, mean alt {firehose.alt_text.str.len().mean():.0f} chars")

See explore.ipynb for a walkthrough with visualizations.

Two Datasets

The targeted corpus and firehose sample are separate files with the same schema. Use them together or independently.

	Targeted corpus	Firehose sample
File	corpus.jsonl.gz	firehose.jsonl.gz
Rows	279,196	125,645
Source	getAuthorFeed from 495 validated accounts	Jetstream live stream (14.5-hour window)
Selection	Accounts screened for high alt-text use; strongly selection-biased	Time-bounded observed Jetstream window
Use case	Training, benchmarking, style analysis	Adoption rates, contrastive training
Reproducibility	Deterministic from account list	Time-dependent snapshot

Schema

Each row is one image with alt text. A post with 4 images produces 4 rows.

Field	Type	Description
alt_text	string	Image description written by the author
image_alt_length	int	Character count of the alt text
text	string	Post body text
author_handle	string	Bluesky handle (e.g. tink.uk)
author_did	string	Decentralized identifier
post_uri	string	AT Protocol post URI
post_cid	string	Content identifier (hash)
created_at	string	Post creation time (ISO 8601)
indexed_at	string	Index time (ISO 8601)
langs_json	string	Language tags as JSON array
image_index	int	Position of this image in the post (0-based)
image_count_in_post	int	Total images in the post
image_mime_type	string	MIME type of the image blob
image_ref	string	AT Protocol blob reference
image_thumb_url	string	CDN thumbnail URL (AppView only)
image_fullsize_url	string	CDN full-size URL (AppView only)
source_mode	string	Collection mode (author_feed or jetstream)
collected_at	string	When this row was collected (ISO 8601)
query	string	Actor handle or search query used
cursor	string	Pagination cursor at collection time
raw_record_json	string	Complete AT Protocol record as JSON

Account Selection

Not a random sample. 495 accounts found through eight strategies, then validated against the live API.

Discovery:

1. Alt Text Hall of Fame honorees and social graph
2. Accessibility keyword search ("alt text", "a11y", "WCAG", "screen reader")
3. Network tracing from 8 accessibility seed accounts
4. International search (German, French, Portuguese, Japanese, Spanish, Dutch, Swedish)
5. Institutional search (museums, zoos, aquariums, space agencies)
6. Photography and visual arts accounts
7. Jetstream firehose sampling for wild-caught writers

Validation gate (every account probed against the live API):

• 90%+ of images have non-empty alt text
• Average alt text 50+ characters
• At least 10 images in sample

Full account list: collector/actors.txt

Known Quirks

• Firehose has no image_thumb_url or image_fullsize_url. Jetstream delivers raw commit records without CDN URLs. These fields are null for firehose rows.
• langs_json is author-declared, not detected. Some posts have empty language tags.
• raw_record_json can be large. Multi-image posts with long text produce records >10 KB. If you don't need provenance auditing, drop this column.
• Quote-posts included. Posts with app.bsky.embed.recordWithMedia (image + quote) are collected alongside standard image posts.

Use Cases

• Fine-tune vision-language models on quality human-written descriptions
• Benchmark generated alt text against real-world examples
• Study how people describe images for screen reader users
• Compare description styles across museums, photographers, scientists, advocates
• Contrastive training: pair corpus (high quality) with firehose (mixed quality)

Distribution

• GitHub: data-poems/bluesky-alt-text
• HuggingFace: lukeslp/bluesky-alt-text

Citation

@dataset{steuber2026blueskyalttext,
  title     = {Bluesky Alt Text Dataset},
  author    = {Steuber, Luke},
  year      = {2026},
  publisher = {GitHub / HuggingFace},
  url       = {https://github.com/data-poems/bluesky-alt-text},
  license   = {See repository rights and limitations notice},
  note      = {279,196 image alt-text pairs from 489 validated Bluesky accounts}
}

Structured Data (JSON-LD)

{
  "@context": "https://schema.org",
  "@type": "Dataset",
  "name": "Bluesky Alt Text Dataset",
  "description": "279,196 image alt-text pairs from 489 validated Bluesky accounts, collected via AT Protocol APIs.",
  "url": "https://github.com/data-poems/bluesky-alt-text",
  "sameAs": [
    "https://huggingface.co/datasets/lukeslp/bluesky-alt-text"
  ],
  "creator": {
    "@type": "Person",
    "name": "Luke Steuber",
    "url": "https://lukesteuber.com"
  },
  "keywords": ["alt text", "accessibility", "image description", "bluesky", "at protocol", "a11y", "screen reader"],
  "temporalCoverage": "2023/2026",
  "distribution": [
    {
      "@type": "DataDownload",
      "encodingFormat": "application/x-ndjson",
      "contentUrl": "https://github.com/data-poems/bluesky-alt-text/releases/latest"
    }
  ]
}

Bluesky Ecosystem

More Bluesky tools by Luke Steuber:

• bsky-firehose-anonymized-dec-2025 -- 101K anonymized firehose posts
• bluedrop -- browser extension for DMs, Zen Mode, and dark mode
• skymarshal -- Python CLI for post and comment management
• skymarshal-js -- TypeScript/JS toolkit for AT Protocol

Rights and limitations

Bluesky users retain ownership of the text they posted. No blanket CC BY grant is claimed for user-authored descriptions, post text, or raw records in these snapshot files. The repository's license file applies only where the repository owner has the rights needed to grant it, such as original code and documentation.

The snapshot was collected from public posts through documented AT Protocol interfaces. “Public” does not mean anonymous: the released schema includes direct and stable identifiers and raw records. Users should account for source authors' rights, deletion requests, contextual integrity, and re-identification risk before reuse.

Author

Luke Steuber

• Website: lukesteuber.com
• Bluesky: @lukesteuber.com
• Contact: luke@lukesteuber.com