This dataset is a raw, unfiltered archive of arguments from the (now defunct) CreateDebate platform. It contains debates on extremely controversial and potentially disturbing topics, including but not limited to: pedophilia, bestiality, rape and sexual violence, suicide and self-harm, hate speech directed at racial, religious, ethnic, and LGBTQ+ groups, graphic violence, and conspiracy/extremist rhetoric. Some arguments express views that the dataset authors find abhorrent. The dataset is released as-is to support academic research in computational argumentation; inclusion of any content is not endorsement.

By requesting access you confirm that:

You are using this dataset for academic or research purposes only.
You will not redistribute the dataset or any substantial portion of it.
You understand the dataset contains content that may be disturbing.
You will not use this dataset to train systems intended to generate,
promote, or amplify the kinds of harmful content it documents.
You will cite the accompanying paper and the Zenodo DOI in any work
derived from this dataset.

CreateDebate Arguments — Full Raw Dump

⚠️ Content warning. This dataset contains user-generated arguments on extremely controversial topics including pedophilia, bestiality, rape, self-harm, hate speech, and graphic violence. It is released unfiltered for research purposes. See the "Considerations for Using the Data" section below. The dataset viewer is disabled by design.

Dataset Description

Dataset Summary

This dataset is the complete raw archive of arguments collected from CreateDebate (createdebate.com, since taken down), comprising 383,220 arguments across the platform's for/against debates. It is released to support reproducible research in computational argumentation, argument mining, stance detection, and online discourse analysis.

A separately released, manually-curated subset of ~84k arguments from 500 debates with sensitive topics removed is available at azza1625/createdebate-arguments. That curated subset, not this one, is the appropriate choice for most downstream applications. This full dump exists for researchers who need the raw distribution — for example to study how argumentative structure varies across topic sensitivity, or to develop content moderation tools.

Supported Tasks

Stance detection — arguments carry an explicit argumentSide label
Argument quality assessment — community vote scores per argument
Argument mining and discourse structure — full tree structure preserved
Computational rhetoric and topic-conditioned analysis
Robustness / safety evaluation — content moderation, toxicity detection

Languages

Primarily English (en). User text is preserved as-written; spelling and grammar quality varies.

Dataset Structure

Files

File	Format	Purpose
`data/train-00000-of-00001.parquet`	Parquet	Full dataset, 383,220 rows
`README.md`	Markdown	This file

Data Fields

Argument-level

Field	Type	Description
`argumentId`	string	Unique argument identifier
`parentId`	int64	Parent argument numeric id; `-1` for top-level
`userPoints`	int64	Lifetime platform reputation of the (anonymized) author
`argumentBody`	string	Full argument text, UTF-8, original whitespace preserved
`date`	int64	Posting time as Unix milliseconds (UTC)
`argumentSide`	string	The side title the argument supports
`argumentPoints`	int64	Net community vote score for this argument
`argumentTag`	string	Optional platform tag (`""` if none)
`characterCount`	int64	Character count of `argumentBody`
`depth`	int64	Tree depth; `0` = top-level, `1` = first reply, …
`lang`	string	Language of `argumentBody`

Debate-level

Field	Type	Description
`debateUrl`	string	Original URL on createdebate.com (no longer resolvable)
`debateTopic`	string	Platform-assigned category
`debateTitle`	string	Debate title
`debateDescription`	string	Debate prompt / description
`debateStats`	string	JSON: `{score, arguments, totalVotes}`
`debateTags`	string	JSON: side → argument count
`debateSides`	string	JSON: list of `{id, title, points}`

Data Splits

No canonical splits are provided. This is a raw distribution; researchers should construct splits appropriate to their task (random, debate-grouped, temporal, or topic-stratified).

Dataset Creation

Source data

Data was scraped from the (now defunct) CreateDebate platform via a Selenium-based crawler. The collection covered all for/against format debates discoverable from the platform's listings during the collection window; debates were prioritized by argument count. For each debate the full argument tree was extracted along with debate-level metadata.

What was filtered

Unlike the curated subset, no content-based filtering was applied to this release. Debates spanning sensitive, offensive, or graphic topics remain in the data. Only the following modifications were made:

All user-identifying columns (username, profileURL) were removed.
parentId values were normalized from the string form "argNNNNN" to the numeric id NNNNN (top-level: -1).
Nested metadata fields were serialized to JSON strings for portability.
Structural fields (depth, context) were computed from the tree.

Annotations

All labels (argumentSide, argumentPoints, debateTopic, …) are user- or platform-generated. No additional human annotation was performed.

Personal and Sensitive Information

All user-identifying fields were removed prior to release. The platform itself has been taken down, so usernames cannot be re-linked from public sources. Argument bodies may, however, mention third parties, public figures, or self-disclose information — they are released as-written.

If you encounter content in this dataset that should be redacted (e.g. personally identifying information that slipped through, content depicting illegal acts against identifiable people), please contact the dataset maintainers to request a redaction in the next version.

Considerations for Using the Data

Intended uses

This dataset is intended for academic research in computational argumentation, NLP, and adjacent fields. Appropriate uses include: linguistic analysis of argumentation, development and evaluation of argument mining systems, stance detection, study of online discourse dynamics, and content moderation research.

Uses we ask you NOT to make

Do not use this dataset to train systems intended to generate, promote, or amplify the kinds of harmful content it documents.
Do not redistribute the dataset or substantial portions of it. Point others to this repository instead.
Do not attempt to re-identify authors.

Biases and limitations

Self-selection bias. CreateDebate users are not representative of any broader population.
Topical skew. Some topics are heavily over-represented; political and ideological debates dominate.
Quality variance. Argument quality ranges from carefully reasoned to trolling, off-topic, or incoherent. Community vote scores partially but imperfectly track quality.
Binary stance simplification. The for/against format flattens nuanced positions.
Harmful content. As stated above, the dataset contains content many readers will find offensive or disturbing.
Temporal drift. Arguments reference contemporary events of their posting period.

Licensing

The compilation, schema, and derived metadata are released under Creative Commons Attribution 4.0 International (CC BY 4.0). The underlying argument text was authored by individual platform users; the platform's terms permitted the collection performed, and the platform is now defunct. Users of this dataset accept the gated-access terms above.

Maintenance

This dataset is a static archive of a defunct platform; the data itself will not grow. Bugfix versions (schema corrections, additional redactions) will be released as new versions of this dataset.

Downloads last month: 12

Total file size:

108 MB