nakas/gpttrace-text-train

GPTTrace text-detection training data

The human and machine text we created or collected to train GPTTrace's AI-text detectors, beyond the public RAID and MAGE benchmarks. It is the redistributable part of the data behind:

• nakas/gpttrace-v5-1b: Ettin-1B, #2 on the RAID leaderboard (99.90% TPR at 5% FPR with adversarial attacks)
• nakas/gpttrace-ettin150-text-v5e: Ettin-150M, #3 on RAID (99.80%), runs in the browser at gpttrace.com

The code that built it (and trains the models) is at github.com/andrewnakas/gpttrace, in eval/train/text/.

What is in it

336,171 texts, label 0 = human, 1 = AI. Columns: text, label, source, generator (the model, or A>B for a rewrite by B of A's text), method (humanizer strategy chain, e.g. wiki_signs+burstiness), licence, attribution, from_dataset.

Config	Rows	Human	AI	Licences
permissive	202,254	103,021	99,233	CC0, public domain, CC BY 4.0, Apache-2.0, PSF-2.0, Europarl notice, Apache/MIT model outputs
share-alike	121,382	74,751	46,631	CC BY-SA 2.5 / 3.0 / 4.0 (Wikipedia, talk pages, Stack Exchange, dolly, MDN, and AI rewrites of them)
cdla-sharing	6,484	4,000	2,484	CDLA-Sharing-1.0 (Topical-Chat and AI rewrites of it)
vendor-outputs	6,051	0	6,051	text written with Claude Opus 5.5 and GPT-6 models (see the notice below)

Human text: Wikipedia, arXiv abstracts, Project Gutenberg, Stack Exchange posts and comments, OpenAssistant oasst2, Databricks dolly, Python / Kubernetes / MDN documentation, Europarl, Ubuntu IRC, Wikipedia talk pages, Topical-Chat, Taskmaster and student essays (ASAP 2.0, PIILO). It also includes a "typo" set of human text with real-world noise.

AI text: written from scratch or rewritten by open models (Qwen3-14B, Qwen3-8B, Qwen2.5-7B, Phi-4-mini, gpt-oss-20b) and by Claude Opus 5.5 / GPT-6. It includes humanizer rewrites: paraphrase, recursive paraphrase, back-translation, burstiness, register shifts, scrubbing "AI tells", avoiding Wikipedia's "signs of AI writing", adding human artifacts (typos, casing), thesaurus swaps and combinations of these. It also includes "mirror" texts, which copy the topic and shape of a specific human text so that only the authorship differs.

Licences and attribution

Each row carries its own licence and attribution. Configs group rows whose licences combine cleanly; follow the terms of the config you use:

• share-alike: adaptations, including AI rewrites of these human texts, remain under CC BY-SA. Give credit as in the attribution column. Texts are excerpted and cleaned (modified).
• cdla-sharing: Topical-Chat © Amazon, CDLA-Sharing-1.0. Keep this config under the same agreement.
• permissive: credit The Learning Agency Lab (ASAP 2.0, PIILO), Google (Taskmaster), The Kubernetes Authors, the Python Software Foundation (© PSF, excerpts of the Python documentation) and "© European Union - Source: European Parliament" (Europarl). arXiv abstracts are CC0 metadata. Gutenberg passages are public domain in the US, with Project Gutenberg headers removed.
• vendor-outputs (notice): written with Anthropic Claude and OpenAI GPT-6 models by the dataset author, who owns the outputs under those services' terms. They are published for research on detecting machine-generated text; your own use may be subject to the respective vendors' terms.

Not included: PubMed Central text (our commercial-subset rows had no per-article licence, so CC BY-ND articles could not be excluded) and the RAID / MAGE benchmark text itself (see below).

RAID and MAGE

The models were also trained on public benchmarks that we do not re-host: RAID (MIT; train split, 500k rows) and MAGE (Apache-2.0; 200k training rows). Load them from their sources. eval/train/text/raid-data/raid_data.py in the GitHub repo reproduces our filtering, including the held-out titles used for our own RAID numbers.

Intended use and limits

For training and evaluating AI-text detectors. The labels mean "written by a person" versus "written or rewritten by a language model". Human texts may contain light machine-assisted editing that we could not detect. Never use a detector's score alone to accuse anyone of using AI.

Citation

@misc{gpttrace-text-train-2026,
  title  = {GPTTrace text-detection training data},
  author = {Nakas, Andrew},
  year   = {2026},
  url    = {https://huggingface.co/datasets/nakas/gpttrace-text-train}
}