SalatielJordao/incels_is_1788055901

Incels.is Public Forum Posts

Sensitive-content warning: This dataset contains raw text from a public incel forum. It may include misogyny, hate speech, harassment, threats, discussions of suicide or self-harm, sexual material, profanity, extremist rhetoric, and other offensive or disturbing content. It may also reproduce personal information that users placed in public posts. Please consider the risks to researchers, annotators, and affected communities.

Dataset summary

This release contains public posts collected from incels.is and enriched with two independent language-identification systems and an English-only AI-text detector. It accompanies and updates the data work described in Multilingual and Informal Web Datasets for Robust Language Modeling.

Last updated: August 29, 2026 (America/Sao_Paulo).
Release identifier: incels_is_1788055901 (Unix timestamp 1788055901).
Data coverage: November 8, 2017 through August 29, 2026.
Rows: 12,950,077 posts from 698,565 collected threads.
Source file SHA-256: 56969d1b4b9d790c226c300866f4d1c9f061e94042f2a3325a03eb98a77a4cd0.

Coverage is not exhaustive

This is not a complete copy of the forum. It contains only threads and posts that were publicly accessible without authentication and successfully reached by the collection pipeline. Private, restricted, deleted, unavailable, or otherwise inaccessible material is excluded. Some public threads or posts may also be absent because of collection scope, historical availability, pagination, deletion, or technical failures. Presence in the source forum does not guarantee presence in this dataset.

Relationship to the paper and update methodology

Collection follows the same general direct-web-scraping methodology described in the paper cited below: public forum listing pages are traversed, public thread pages are downloaded, and individual posts and thread metadata are normalized into tabular rows. For this 2026 update, public forum listings and thread pages were revisited so that reachable newer threads and additional pages of previously archived threads could be incorporated. Duplicate comment identifiers were checked during assembly.

The language and AI annotations in this release extend the original collection methodology. They are machine-generated analytical fields and are not human ground-truth annotations.

Files and loading

The default configuration contains one Parquet file in data/. Empty-text rows are intentionally retained.

from datasets import load_dataset

dataset = load_dataset(
    "SalatielJordao/incels_is_1788055901",
    split="train",
    token=True,
)

Access requires a Hugging Face account, acceptance of the conditions above, and manual approval by the dataset author.

Data fields

Field	Description
site	Source site (incels.is).
thread_num	Numeric source thread identifier.
thread_url	Public source thread URL. Links may later stop working.
forum_name	Public forum/section name.
thread_title	Thread title at collection time.
thread_label	Source label such as Serious, JFL, or Blackpill; empty when unlabeled.
comment_id	Unique composite thread/post identifier.
author	Public display name at collection time; not a verified identity.
timestamp	Original displayed timestamp.
timestamp_utc	Normalized timezone-aware UTC timestamp.
timestamp_epoch_ms	Timestamp in Unix milliseconds.
response_id	Post number within its thread.
response_id_reference	Referenced post number, or -1 when none was extracted.
response	Extracted public post text; may be empty for posts with no extractable text.
detected_language_id	Numeric class from the fine-tuned byte-level language classifier.
detected_language	Its NLLB-style language/script prediction, e.g. eng_Latn.
detected_language_confidence	Its top-1 confidence.
ai_generated_probability	Desklib AI-text probability for eligible English rows.
ai_generated	Thresholded AI-text decision or the documented pre-cutoff assumption.
ai_detection_basis	Distinguishes model-scored rows from pre-GPT-2 assumed-human rows.
fasttext_language	Native top-1 label from fastText lid.176.bin, e.g. en.
fasttext_language_confidence	fastText top-1 probability.

Models and annotation procedures

Fine-tuned byte-level language classifier

The detected_language* fields were produced by a compact fine-tuned byte-level language classifier with 64 language/script classes. The checkpoint and model card are planned for a separate Hugging Face release. Empty or whitespace-only responses receive null language fields.

fastText baseline

The fasttext_language* fields were produced with Meta/Facebook Research's full lid.176.bin model, which predicts among 176 languages. The model was downloaded from the official fastText release and pinned locally with SHA-256 7e69ec5451bc261cc7844e49e4792a85d7f09c06789ec800fc4a44aec362764e. Embedded line boundaries were replaced with spaces for one-line fastText prediction. Empty responses receive null fastText fields.

• Model documentation: https://fasttext.cc/docs/en/language-identification
• fastText code: https://github.com/facebookresearch/fastText

AI-text detector

Only responses predicted as English by the fine-tuned byte-level language classifier were eligible for AI-text annotation. English messages before 2019-02-14T00:00:00Z, GPT-2 publication date, are marked assumed-human with ai_detection_basis=assumed_human_pre_gpt2 and have no model probability. English messages at or after the cutoff were scored with desklib/ai-text-detector-v1.01, pinned at revision 5fdea974cd4287c61674951ec78803aa274e2fb7, truncated to 768 tokens, and thresholded at probability >= 0.5. Non-English and empty responses have null AI fields.

Critical accuracy and short-text disclaimer

All language and AI fields are model predictions, not verified facts. Short responses are particularly unreliable:

• agreement between the two language classifiers was only 51.67% below 20 characters, compared with approximately 94–96% for most bands of 50 or more characters;
• the fine-tuned language classifier sometimes assigns incorrect non-English labels to short English replies with high confidence;
• fastText was trained on cleaner sources and can also misclassify informal or code-switched text;
• the AI detector labeled 85.61% of eligible texts below 20 characters and 72.67% of texts from 20–49 characters as AI, showing severe length sensitivity in this domain.

The two language classifiers agree on 84.38% of all nonempty rows, but this is agreement, not accuracy. The AI fields should not be used to accuse or make claims about individual authors.

Dataset statistics

• Total rows: 12,950,077
• Distinct collected threads: 698,565
• Nonempty responses classified by both language systems: 11,549,387
• Empty responses retained with null language fields: 1,400,690
• Rows predicted as English by the fine-tuned classifier: 10,011,840
• Post-cutoff English rows scored by the AI detector: 8,475,923
• Pre-cutoff English rows assigned the assumed-human basis: 1,535,917

Intended uses

Appropriate uses include controlled academic research in computational social science, sociolinguistics, online-community studies, content moderation, language identification, robustness, and dataset-quality analysis.

Prohibited and discouraged uses

Access does not authorize users to:

• identify, deanonymize, contact, surveil, or harass forum participants;
• operationally profile or target individuals or groups;
• use the dataset in violation of applicable law, institutional policy, source rights, or Hugging Face policies.

Privacy, copyright, and removal requests

The dataset contains material that was publicly visible at collection time, including display names and direct public URLs. Public availability does not eliminate ethical or legal obligations. The source site's terms state that post authors retain copyright over their content. This release therefore uses license: other; access to the dataset does not transfer ownership of source posts or grant rights beyond the approved research use.

Researchers should avoid reproducing raw posts. Requests concerning correction, removal, or access should be submitted to the dataset owner through the Hugging Face repository's Community tab or the contact information associated with the paper. Applicable requests will be reviewed for future versions, but removal from this research copy cannot remove content from the source forum or copies already obtained by approved users.

Limitations and biases

• The corpus represents one particular online community and must not be used to generalize about men, involuntary celibates, or broader populations.
• Collection is incomplete and availability-based; missingness is not random.
• Deleted, restricted, inaccessible, and uncollected posts are absent.
• Moderation, bans, deletions, forum structure, and scraper behavior affect the observed distribution.
• Text extraction may omit media, formatting, reactions, edits, or quoted context; 1,400,690 retained rows have empty extracted response text.
• Language confidence is not calibrated ground truth, especially for short or mixed-language posts.
• AI-text predictions are strongly length-dependent and should not be treated as estimates of true AI prevalence.

Citation

If you use this dataset, please cite the associated paper:

Salatiel A. A. Jordão, Carine S. Santos, and Maria J. P. Dantas, "Multilingual and Informal Web Datasets for Robust Language Modeling," in 2025 IEEE 49th Annual Computers, Software, and Applications Conference (COMPSAC), 2025, pp. 1534–1535, doi: 10.1109/COMPSAC65507.2025.00199.

@inproceedings{jordao2025multilingual,
  author    = {Jord{\~a}o, Salatiel A. A. and Santos, Carine S. and Dantas, Maria J. P.},
  title     = {Multilingual and Informal Web Datasets for Robust Language Modeling},
  booktitle = {2025 IEEE 49th Annual Computers, Software, and Applications Conference (COMPSAC)},
  year      = {2025},
  pages     = {1534--1535},
  doi       = {10.1109/COMPSAC65507.2025.00199}
}