codex dataset card

> generated by Codex_

Dataset Card for lesswrong_260509

Dataset Summary

A structured crawl of public longform posts and comments from LessWrong, a ForumMagnum-powered discussion site focused on rationality, cognitive science, AI alignment, forecasting, philosophy, and related topics. The dataset contains 45,933 posts and 906,338 comments, preserving one flat comment list per post with parent IDs, nesting depth, orphan markers, timestamps, scores, vote counts, author metadata, and post-level flags.

The crawl targets public, approved, non-draft, non-future, listed, non-shortform posts visible to an anonymous LessWrong GraphQL requester.

---

Dataset Details

Dataset Sources

• Source Website: https://www.lesswrong.com
• Source API: https://www.lesswrong.com/graphql

---

Uses

Direct Use

Suitable for:

• Academic research into forum dynamics and longform online discussion
• Computational social science research on rationalist and AI safety communities
• Linguistic analysis of argumentative, technical, and speculative web discourse
• Studying thread/reply structures, voting patterns, and moderation markers
• Information retrieval and semantic search over LessWrong posts and discussions

---

Dataset Structure

Schema

Column	Type	Description
source	struct	Crawl/source metadata for the row
post	struct	Post metadata and body content
comments_flat_count	int64	Number of comments in the comments list
comments	list	Flat list of comments associated with the post

Source Struct

Field	Type	Description
lesswrong_graphql	string	LessWrong GraphQL endpoint used for the crawl
lesswrong_url	string	Direct LessWrong URL for the post
greaterwrong_url	string	Corresponding GreaterWrong mirror URL
fetched_at	string	ISO-8601 timestamp when the record was fetched

Post Struct

Field	Type	Description
_id	string	Unique post identifier
title	string	Post title
slug	string	URL slug
pageUrl	string	Full LessWrong URL
pageUrlRelative	string	Relative LessWrong URL path
url	string or null	External URL for link posts
postedAt	string	Post timestamp (ISO-8601)
baseScore	int64	Net score at crawl time
voteCount	int64	Vote count at crawl time
wordCount	int64	Word count reported by LessWrong
htmlBody	string	Post body in HTML
meta	bool	Whether the post is marked as meta
question	bool	Whether the post is marked as a question
isEvent	bool	Whether the post is an event post
frontpageDate	string or null	Frontpage timestamp, if present
curatedDate	string or null	Curated timestamp, if present
legacyId	string or null	Legacy LessWrong identifier, if present
user	struct	Author info (_id, username, slug, displayName)

Comment Struct

Field	Type	Description
_id	string	Unique comment identifier
postId	string	Parent post ID
parentCommentId	string or null	ID of the parent comment this replies to
topLevelCommentId	string or null	ID of the top-level ancestor comment
pageUrl	string	Direct LessWrong URL for the comment
pageUrlRelative	string	Relative LessWrong URL path
postedAt	string	Comment timestamp (ISO-8601)
baseScore	int64	Net score at crawl time
voteCount	int64	Vote count at crawl time
wordCount	int64	Word count reported by LessWrong
htmlBody	string	Comment body in HTML
deleted	bool	Whether the comment is deleted
deletedPublic	bool	Whether the deletion is public
retracted	bool	Whether the comment is retracted
legacyId	string or null	Legacy LessWrong identifier, if present
user	struct	Author info (_id, username, slug, displayName)
position	int64	1-based position in pre-order traversal within the post's comment list
depth	int64	Nesting depth (0 = top-level comment or orphan-root comment)
orphan	bool	Whether the comment was associated with the post but not attached to the main comment tree

Statistics

Metric	Value
Posts	45,933
Non-orphan comments	902,220
Orphan comments	4,118
Total comments	906,338
Post date range	2005-01-01 to 2026-05-06
Comment date range	2006-11-22 to 2026-05-09
Crawl timestamp range	2026-05-06 to 2026-05-09
Mean comments per post	19.7
Median comments per post	4
Longest post thread	2,001 comments
Max nesting depth	91
Post authors	6,581
Comment authors	22,139
Posts with no author	1,510 (3.3%)
Link posts	5,516 (12.0%)
Text posts	40,417 (88.0%)
Posts with no comments	12,059 (26.3%)
Meta posts	213 (0.5%)
Question posts	2,689 (5.9%)
Event posts	2,095 (4.6%)
Frontpage posts	22,581 (49.2%)
Curated posts	738 (1.6%)
Negative-score posts	1,904 (4.1%)
Mean post word count	1,324
Median post word count	577
Mean post body HTML length	19,154 characters
Median post body HTML length	4,742 characters
Mean comment word count	106
Median comment word count	60
Blank comment bodies	2,336 (0.3%)
Deleted comments	1,838 (0.2%)
Retracted comments	10,239 (1.1%)

The most prolific post authors include Eliezer Yudkowsky (1,048 posts), Zvi (989), jefftk (921), Stuart_Armstrong (811), and KatjaGrace (632). The most prolific comment authors include ChristianKl (13,987 comments), wedrifid (13,141), Lumifer (11,480), gwern (10,952), and Vladimir_Nesov (8,612).

---

Dataset Creation

Curation Rationale

LessWrong is a long-running forum for rationality, AI alignment, decision theory, philosophy, cognitive science, and adjacent technical discussion. A structured archive makes this corpus available for research into longform online discourse, intellectual community dynamics, and the development of AI safety and rationalist discussion over time.

Source Data

Data Collection and Processing

• Tool: Custom Python script using the public LessWrong GraphQL API
• Scope: Public, approved, non-draft, non-future, listed, non-shortform posts visible to anonymous GraphQL requests
• No filtering was applied to fetched content; all crawled posts and comments are included
• Comments are stored as a single flat list in the comments column
• Parent-child relationships are preserved using parentCommentId, topLevelCommentId, position, and depth
• Comments that could not be attached to the main comment tree are retained in comments with orphan = true
• Score and vote counts are snapshot values from crawl time

Who are the source data producers?

LessWrong users posting under real names or pseudonymous usernames.

---

Bias, Risks, and Limitations

Content Risks

• Content reflects the interests, norms, and epistemic commitments of the LessWrong community
• Posts may contain speculative, controversial, or technical claims presented with varying levels of evidence
• Discussions may include sensitive topics such as AI existential risk, mental health, bioethics, politics, religion, self-experimentation, and personal life decisions
• Historical posts may contain outdated claims, broken links, or references to community controversies

Structural Limitations

• The crawl was conducted from 2026-05-06 to 2026-05-09 and includes no content discovered beyond that snapshot
• The latest post in the shards is dated 2026-05-06, while the latest comment is dated 2026-05-09
• Vote counts, scores, curation status, frontpage status, deletion markers, and retraction markers are snapshot values and may have changed since crawl time
• The dataset is limited to content discoverable through the public GraphQL API without privileged access
• Some older imported LessWrong content has missing author metadata or legacy formatting artifacts in htmlBody
• A small number of comments are marked with orphan = true because they could not be attached to the main comment tree

Personal and Sensitive Information

• No systematic PII detection or redaction was performed
• Many users post under stable pseudonyms, and some post under real names
• Posts and comments may contain references to real individuals, organizations, events, or personal experiences
• The dataset should be treated as containing potentially sensitive information

Licence and Redistribution

No licence was identified for the source forum content. This dataset is therefore marked as license: unknown.

The dataset consists of scraped public forum posts authored by third parties. The dataset publisher does not claim ownership of the underlying post text and does not grant additional rights beyond any rights users may already have under applicable law.

Users are responsible for determining whether their intended use is lawful, including with respect to copyright, privacy, data protection, and platform terms.