SalatielJordao/orkut-communities

Orkut Communities Archive Threads

Dataset Summary

This dataset contains text extracted from archived Orkut community pages. It is derived from Internet Archive / Archive Team Orkut WARC and CDX files and converted into a hierarchical Parquet dataset for research use.

The dataset focuses on human-authored text fields found in community and forum pages, including community names and descriptions, forum topic metadata, usernames, reply dates, reply titles, and reply bodies. Images and other binary assets are not included.

Each row represents one forum topic/thread. Replies are stored as a nested list inside their parent thread row.

Dataset Structure

Files

The release is stored as native Parquet shards:

train-00000-of-00151.parquet
...
train-00150-of-00151.parquet
_build_manifest.json

Split

Split	Rows	Description
train	120,640,699	One row per Orkut community forum thread/topic

The nested replies field contains 897,280,968 replies according to the generated reply_count values.

Data Fields

Top-level thread fields:

• community_id: Orkut community identifier.
• community_name: Extracted community name.
• community_description: Extracted community description when available.
• community_url: Archived Orkut community URL.
• community_timestamp: Archive capture timestamp for the selected community metadata.
• community_info: List of key/value metadata items extracted from the community profile page.
• topic_id: Forum topic/thread identifier.
• topic_title: Topic title.
• topic_href: Topic link as found in forum listings.
• topic_url: Archived topic URL when available.
• topic_timestamp: Archive capture timestamp for the topic page.
• summary: Topic summary text when available.
• removed_message_count: Count of removed messages detected on the topic page when available.
• reply_count_text: Reply count text from forum listings, preserved as extracted.
• latest_excerpt: Latest-reply excerpt from forum listings when available.
• latest_date: Latest-reply date text from forum listings when available.
• forum_index_refs: List of forum index pages where the topic appeared.
• reply_count: Number of nested reply records in replies.
• replies: Nested list of reply structs.

Nested reply fields:

• reply_id: Reply identifier from the archived URL.
• author_id: Author identifier from the archived URL when available.
• author_name: Display name extracted from the reply page.
• date: Original reply date string as extracted from Orkut HTML.
• date_iso: Normalized date in YYYY-MM-DD format when parseable.
• reply_title: Reply-page title/subject text.
• body: Reply body text.
• url: Archived reply URL.
• timestamp: Archive capture timestamp.
• digest: WARC/CDX digest.
• warc: Source WARC path within the local Archive Team item layout.
• offset: Source WARC offset.
• size: Source WARC record size.

Loading

Streaming is recommended because the dataset is large.

from datasets import load_dataset

ds = load_dataset("SalatielJordao/ORKUT_IA_communities_text", split="train", streaming=True)
row = next(iter(ds))
print(row.keys())

With Polars:

import polars as pl

lf = pl.scan_parquet("/path/to/train-*.parquet")
summary = lf.select([
    pl.len().alias("threads"),
    pl.col("reply_count").sum().alias("replies"),
]).collect()
print(summary)

Dataset Creation

Source Data

The source data consists of archived Orkut pages captured by Archive Team and preserved by the Internet Archive. The conversion process used CDX indexes and WARC records to extract HTML pages matching Orkut community, forum index, topic, and reply URL patterns.

Extraction and Processing

The extraction pipeline selected text-oriented HTML pages and ignored images and other binary resources. It extracted four source record categories:

• community: Community profile pages.
• forum_index: Forum listing pages containing topic lists and latest-reply snippets.
• topic: Topic/thread pages.
• reply: Individual reply pages.

The final Parquet builder groups records by (community_id, topic_id), creates one row per thread, nests replies inside that row, and writes Hugging Face-style Parquet shards. Replies are ordered by date_iso, then reply_id, then archive timestamp and URL as tie-breakers.

Exact duplicate reply records may be removed during conversion when they have the same (reply_id, author_id, url, digest) within a thread. This deduplication is record-level only; repeated or copied human text is preserved.

Known Caveats and Limitations

• The dataset contains archived user-generated content. Text may include offensive, adult, hateful, political, personal, or otherwise sensitive material.
• User display names, author identifiers, URLs, and message text are preserved when extracted.
• The original Orkut pages were archived from the web; archival availability does not necessarily imply that every item is appropriate for every downstream use.
• Some archived pages contain encoding damage or mojibake. In a small number of replies, date_iso may be empty even though the original date string is present.
• Some threads have zero nested replies because they appeared in forum index pages but no reply page was extracted for that topic.
• Archive capture timestamps are not the same as message creation times.
• The dataset is not a complete reconstruction of Orkut. It reflects what was present in the source WARC/CDX files and what the extractor could parse.
• No images, profile photos, attachments, or non-text resources are included.

Ethical, Legal, and Privacy Considerations

This dataset includes historical social-media/forum content and may contain personal data, usernames, identifiers, and self-disclosed information. Users of this dataset are responsible for determining whether their use complies with applicable law, platform policy, institutional review requirements, and ethical research standards.

Recommended uses include historical web research, computational social science, language modeling research, web-archive analysis, and dataset preservation studies.

Discouraged uses include deanonymization, doxxing, targeted profiling, harassment, surveillance, attempts to contact or identify individual users, or making consequential decisions about people represented in the data.

If you are a rights holder or an affected individual and want specific content removed from a downstream hosted copy, contact the dataset maintainer through the Hugging Face repository discussion or issue mechanism.

Licensing Information

The extracted data is derived from archived Orkut user-generated content and associated archived pages. No clear machine-readable open license was identified for the original user-authored content. For that reason, the dataset metadata uses license: other.

Users must make their own assessment of copyright, privacy, data protection, and other legal obligations before redistributing or using this dataset.

Citation

If you use this dataset, please cite the Hugging Face dataset repository and acknowledge the Internet Archive / Archive Team source materials.

Suggested citation:

@dataset{orkut_communities_archive_threads,
  title = {Orkut Communities Archive Threads},
  year = {2026},
  publisher = {Hugging Face},
  note = {Derived from Internet Archive / Archive Team Orkut web archives}
}

Build Metadata

A machine-readable build summary is included in _build_manifest.json. At the time this card was written, the build manifest reported:

• Thread rows: 120,640,699
• Parquet shards: 151
• Temporary bucket count used during conversion: 2,048
• Compression: Zstandard
• Compression level: 6
• Target shard size: 1,024 MB
• Target row group size: 256 MB

---

Dataset card written by Codex (GPT 5.5 XHigh) on May 21, 2026.