README.md

---
license: cc-by-nc-sa-4.0
language:
- en
pretty_name: WildAI
task_categories:
- text-generation
tags:
- ai-generated-text
- pretraining
- common-crawl
configs:
- config_name: human
  data_files:
  - split: train
    path: human/*.parquet
  default: true
- config_name: ai
  data_files:
  - split: train
    path: ai/*.parquet
- config_name: mixed
  data_files:
  - split: train
    path: mixed/*.parquet
- config_name: labels
  data_files:
  - split: train
    path: labels/*.parquet
---

# WildAI

> **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.

WildAI is a pool of English web documents from Common Crawl (2024 to mid-2026), each labeled by the Pangram AI-text
detector as human-written, AI-generated or mixed. It was assembled to build pretraining mixtures with a controlled share
of AI-generated text for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295).

**WildAI is not a natural sample of the web.** An EditLens detector preselected likely-AI documents, so AI text is far
more common in the pool than on the web; `selection` records why each document is in it.

## Composition

| Config | Source | Documents | GPT-2 tokens |
|---|---|---:|---:|
| `human` | common_crawl | 4,177,080 | 2,640,869,930 |
| `human` | fineweb | 54,736,995 | 37,190,059,660 |
| `ai` | common_crawl | 1,490,771 | 1,677,945,232 |
| `ai` | fineweb | 31,050,180 | 32,198,690,948 |
| `mixed` | common_crawl | 262,131 | 312,479,785 |
| `mixed` | fineweb | 4,325,289 | 5,268,906,826 |

## Configs

| Config | Contents |
|---|---|
| `human`, `ai`, `mixed` | Pool documents by Pangram label. |
| `labels` | Every pool document's metadata and labels, without text. |

## How it was built

1. **Collection.** Documents from the Hugging Face FineWeb v1.4.0 dumps CC-MAIN-2024-10 to CC-MAIN-2025-26 (`source =
   fineweb`), and from WARC files of the crawls CC-MAIN-2025-30 to CC-MAIN-2026-25 processed with FineWeb's recipe in
   DataTrove 0.2.0 (`source = common_crawl`). These crawls were processed from the first 2.6 % of each crawl's WARC files, without FineWeb's MinHash deduplication.
2. **Preselection.** EditLens (`pangram/editlens_Llama-3.2-3B`, four buckets, mean over up to three 512-token windows)
   labeled the collected documents. Documents in its two AI buckets and a crawl-matched draw from its human bucket
   became candidates (`selection` = `editlens_ai`, `editlens_human`); its lightly-edited bucket was dropped. A
   label-blind random draw from the 2026 crawls was labeled too (`selection = natural`).
3. **Labeling.** Pangram labeled every candidate as Human, Mixed or AI (`pangram_label`), with the fractions of the text
   classified AI, AI-assisted and human. Pangram 3.3.2 scored up to three disjoint 512-token windows of each document (the first, middle and last): a document is Human or AI only when every window agrees, and Mixed otherwise, and its fractions are the shares of its windows.
4. **Topics and formats.** WebOrganizer's topic and format classifiers (`WebOrganizer/TopicClassifier`,
   `WebOrganizer/FormatClassifier`) labeled each document.
5. **Pools.** Documents were split by Pangram label and deduplicated by `id`.

## Fields

`id` is the Common Crawl WARC-Record-ID. `token_count` counts GPT-2 tokens of the released text. `truncated` marks text
shorter than the page's extraction. `warc_path` is empty for documents taken from Hugging Face FineWeb (its `file_path`,
joined on `id`, gives it). Sorting a pool by `sampling_hash` gives its order: any prefix is a uniform sample. `pii_anonymized_after_labeling` marks rows whose e-mail or IP addresses were replaced
at release with FineWeb's placeholders, after the labels were computed on the original text.

## Limitations

- Detector labels are not ground truth: Pangram makes errors, and Mixed documents are ambiguous.
- The pool over-represents AI text by design and covers only the crawl files that were processed.
- Deduplication is by document id; the same text can occur under different ids, across crawls.

## Personal information and removal requests

Text went through FineWeb's anonymization of e-mail and IP addresses; other personal information may remain. To request
removal of content, contact the dataset maintainers through the Hugging Face discussion tab.

## License

**License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license. The full text is in `LICENSE`. The text comes from FineWeb (ODC-By 1.0) and Common Crawl, and remains subject to Common Crawl's terms of use.

## Citation

```bibtex
@article{russell2026wildai,
  title   = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
  author  = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
  journal = {arXiv preprint arXiv:2609.40295},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.40295}
}
```