Datasets:

rishanthrajendhran
/

IdeaShift-X

license: cc-by-nc-sa-4.0
language:
  - am
  - ar
  - bn
  - de
  - es
  - fa
  - fr
  - hi
  - id
  - ja
  - km
  - ml
  - mr
  - ne
  - ru
  - si
  - sw
  - ta
  - th
  - uk
  - ur
  - vi
  - yo
  - zh
tags:
  - ai-text-detection
  - idea-provenance
  - multilingual
pretty_name: IdeaShift-X
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: source
        path: data/source.parquet
      - split: level_0
        path: data/level_0.parquet
      - split: level_5
        path: data/level_5.parquet

IdeaShift-X

IdeaShift-X tests idea-provenance detection outside English. It holds about 10,000 human-written web pages in 24 languages from FineWeb2. Each page has two AI versions in the same language, written by Gemini 3.7 Flash:

Level 0: written from a brief naming only the document type, so the ideas are the AI's.
Level 5: written from the page's full role-labelled outline with an instruction to add nothing, so the ideas are the page's and the prose is the AI's.

A detector should not flag the human pages. It should flag level 0 for AI ideas. An idea detector should not flag level 5; a prose detector should.

Construction

The dataset merges two builds.

9-language build: Arabic, Chinese, German, Indonesian, Japanese, Russian, Spanish, Tamil, Urdu. About 610-740 pages per language.
24-language build: adds 15 languages at about 250 pages each.

Pages were format-classified with WebOrganizer's annotation prompt on an LLM. Only pages in the eight long-form formats IdeaLens was trained on were kept, some of them force-fitted to the closest format. Outlines are always extracted in English, whatever the page's language.

Splits

Split	Rows	What it holds
`source`	10,160	the human-written FineWeb2 pages
`level_0`	10,152	the AI version written from a brief naming only the document type (AI ideas)
`level_5`	10,116	the AI version written from the page's full outline (human ideas, AI prose)

Rows in different splits share seed_id; a few pages lack one of their AI versions. language selects a language, e.g. load_dataset("rishanthrajendhran/IdeaShift-X", split="level_5").filter(lambda r: r["language"] == "fra_Latn").

Fields

Field	Description
`id`	`<language>#<FineWeb2 id>` for human pages; `…\|level0` / `…\|level5` for AI versions
`seed_id`	the human page the row belongs to
`language`, `language_name`	FineWeb2 language code and name
`build`	`9-language` or `24-language`
`level`	null for the human page; 0 or 5 for AI versions
`kind`, `prose_author`, `idea_author`	what the row is and who wrote its prose and ideas
`generator`	`gemini-3.7-flash` for AI versions
`format`	the page's format (WebOrganizer taxonomy, one of the eight long-form formats)
`words`, `text`	the document
`seed_crawl_dump`, `seed_crawl_date`	Common Crawl snapshot and crawl date of the human page
`system_prompt`, `prompt`	the exact generation request of a level-0 document (`system_prompt` is always "You are a skilled writer."); empty for the human pages and for level 5
`outline`	the English, role-labelled outline of this document as IdeaLens scored it (JSON string: `document_description`, `global_themes`, `items` of `role` and `content`)
`idealens_p_ai`, `proselens_p_ai`, `pangram4_p_ai`	detector P(AI) as reported in the paper (IdeaLens on the outline, ProseLens and Pangram 4 on the document)

Caveats

URLs are not included. The FineWeb2 id identifies each page in the source dataset.
Outlines are the untreated outlines each document was scored on, not the paraphrased outlines used for training. The level-5 prompt is assembled from a template for each document and was not stored, so system_prompt and prompt are empty there.
Scores are stored at full precision. Rounding them ties the very small P(AI) values and moves the AUC. pangram4_p_ai is empty for a few documents Pangram returned nothing for.
Flag a document at the 1% global cut, as in the paper: IdeaLens when idealens_p_ai is above 0.86288, Pangram 4 when pangram4_p_ai is at least 0.5, ProseLens when proselens_p_ai is above 0.39461.
Outlines and page text can contain web addresses or e-mail addresses that were on the original page.

Scoring

The idealens package (PyPI) runs IdeaLens, ProseLens and the other released detectors on these documents: pip install idealens, then idealens run on a JSONL file with a text field.

All IdeaLens models and datasets are in the IdeaLens collection.

License

The dataset is released under CC BY-NC-SA 4.0: free to share and adapt for non-commercial purposes, with attribution and under the same license. The human pages come from FineWeb2 (ODC-By 1.0) and remain subject to the terms of their original sources and of Common Crawl. The AI versions were produced with Google Gemini models.

Citation

@article{idealens2026,
  title         = {IdeaLens: Detecting AI Ideas in Long-form Writing},
  author        = {Rajendhran, Rishanth and Choi, Minjoon and Russell, Jenna and Namuduri, Ramya and B{\"o}l{\"o}ni-Turgut, Deniz and Karpinska, Marzena and Wieting, John and Iyyer, Mohit},
  journal       = {arXiv preprint arXiv:2610.06778},
  year          = {2026},
  eprint        = {2610.06778},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2610.06778}
}