Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    ValueError
Message:      Bad split: corpus. Available splits: ['train']
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 99, in get_rows_or_raise
                  return get_rows(
                         ^^^^^^^^^
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                         ^^^^^^^^^^^^^^^^^^^^^
                File "/src/services/worker/src/worker/utils.py", line 61, in get_rows
                  ds = load_dataset(
                       ^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/load.py", line 1705, in load_dataset
                  return builder_instance.as_streaming_dataset(split=split)
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 1196, in as_streaming_dataset
                  raise ValueError(f"Bad split: {split}. Available splits: {list(splits_generators)}")
              ValueError: Bad split: corpus. Available splits: ['train']

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Bluesky Alt Text: 279K Curated Image Descriptions

License: CC BY 4.0 HuggingFace GitHub Rows Authors

279,196 image alt-text pairs from 489 Bluesky accounts, collected via the public AT Protocol APIs. Every account validated at 90%+ alt-text rate with substantive descriptions. One row per image, not per 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):

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)
Alt text quality High (accounts verified at 90%+ rate) Mixed (reflects platform-wide behavior)
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

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   = {CC-BY-4.0},
  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"
  ],
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "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:

License

CC-BY 4.0. Use it for anything, just credit the source.

Collected from public Bluesky posts via documented AT Protocol APIs. No private data, no scraping, no authentication required.

Author

Luke Steuber

Downloads last month
55