WildOutlines

WildOutlines is the training corpus of IdeaLens, a detector of who came up with the ideas in a document, rather than who wrote its words. It pairs 1,001,145 English web documents with role-labelled outlines of their ideas: an ordered list of items, each giving one idea and the role it plays in the document (for example Central Development, Source Viewpoint, Open Question).

Every document carries a human or AI label. Each outline comes in two versions: the outline as extracted, which is what IdeaLens reads when scoring a new document, and a paraphrased outline with the document's wording removed, which is what IdeaLens was trained on.

Quick start

from datasets import load_dataset

train = load_dataset("rishanthrajendhran/WildOutlines", split="train")
row = train[0]
print(row["label"], row["format"])
for item in row["paraphrased_outline"]["items"]:
    print(f"[{item['role']}] {item['content']}")

IdeaLens reads an outline as one line per item, [role] content, joined with newlines.

Splits

calibration contains only human documents, 10,000 per format, and is disjoint from validation. The IdeaLens decision thresholds are the score levels that flag 1% of these documents, either over all of them (global cut) or within each format (per-format cut). validation is mostly human, so report AUC or balanced metrics on it rather than accuracy. test is close to balanced within every format.

Fields

Both outline fields share one structure:

document_description: string          # what kind of document this is: genre, audience, register
global_themes:        list<string>    # the document's main points
items: list<{
    role:     string,                 # the role this item plays, from the format's role vocabulary
    content:  string,                 # one idea, stated in one or two sentences
    verbatim: bool                    # true if the content quotes the document rather than summarising it
}>

The paraphrase keeps every item's role, order and content, and changes only the wording.

How the corpus was built

• Documents and labels. Documents are English Common Crawl pages from the 2024, 2025 and 2026 crawls, drawn from the WildAI corpus (Russell et al., 2026), which builds on FineWeb and annotates each page with a WebOrganizer format and topic and a Pangram authorship prediction. We kept documents of at least 500 words, in eight long-form formats, that Pangram labels as entirely human-written or entirely AI-generated. Documents Pangram labels as mixed or AI-assisted were left out. The paper describes the construction in full.
• Outline extraction. For each format, a role vocabulary lists the recurring functions an idea can serve in that kind of document. An LLM, given the document, its format's role vocabulary and six worked examples, writes the outline. Most documents were extracted by gemini-3.7-flash; part of the corpus, from an earlier build, by gemini-3.1-pro. The extractor field records which.
• Paraphrasing. gemini-3.1-pro rewrote every outline item without seeing the source document, so that the document's own wording does not carry into training.

The extraction prompts, role vocabularies and worked examples ship with the idealens package (PyPI), which runs the same extraction on new documents and scores them with the released detectors.

Labels

The labels come from the data source, which assigned them with the Pangram AI-text detector: Pangram 3.3.2 for pages from the 2024 and 2025 crawls, and Pangram 4 for pages from the 2026 crawls (metadata.pangram.model records which). They are silver labels: they record whether Pangram judged the prose to be human-written or AI-generated, not an annotation of who conceived the ideas. IdeaLens is trained on these prose labels through outlines whose wording has been removed, which pushes it to learn from the ideas themselves.

Metadata

metadata holds the detector outputs that exist for a document. Each block is either complete or null:

metadata:
  pangram:                           # every document
    prediction:       "Human" | "AI"  # the verdict the label was taken from
    model:            "pangram-3" | "pangram-4"  # the Pangram model behind the label
    version:          string | null   # "4.0" on the 10,000 documents we scored with Pangram 4 ourselves
    score_confidence: {score, confidence} | null      # 603,132 documents
    fractions:        {ai, ai_assisted, human} | null # 97,675 documents (2026 crawls): share of the text in each class
  editlens:           {score, bucket} | null          # 57,886 documents (2026 crawls), EditLens AI-editing score

Intended use

Training and evaluating detectors of idea provenance, and studying how human and AI-generated documents differ in their ideas and structure. A detector trained on this corpus estimates the provenance of a document's ideas; it should not be the sole basis for decisions about a person's work.

Limitations

• English only; documents of at least 500 words; eight long-form web formats.
• Labels are detector-derived (see Labels).
• Two extraction models were used; detectors trained here are expected to be robust to the extractor, and the paper tests this.
• Creative Writing and Personal About Page lean human (few AI-labelled documents were available in these formats).
• Outlines were written by LLMs and can contain extraction errors.

Related

All IdeaLens models and datasets are in the IdeaLens collection. The idealens package runs them.

License

The dataset (outlines, paraphrases, labels and metadata, and its compilation) 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 document text comes from FineWeb (ODC-By 1.0) via WildAI and remains subject to the terms of its original sources and of Common Crawl. The outlines were generated 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}
}