openai/gpt-2-output-dataset
OfficialDataset of GPT-2 outputs for research in detection, biases, and more
openai/gpt-2-output-dataset is OpenAI's official dataset of GPT-2 model outputs, released under the MIT license to support machine-generated-text detection research. It's the archived, still-downloadable dataset researchers cite when benchmarking AI-text detectors against GPT-2-era writing.
Our Review
OpenAI released this dataset in 2019 as the companion data for GPT-2's staged rollout, and it remains the reference training set researchers cite when building machine-generated-text detectors. The repo carries 2,026 GitHub stars and 546 forks, and although GitHub marked it archived, the underlying data is still hosted and downloadable today.
Key capabilities:
- •750K labeled documents per model size 250K WebText human-written test docs, 250K random-sampled GPT-2 generations, and 250K Top-K 40 truncated generations, each split into train, validation, and test (250K/5K/5K).
- •Coverage across all four GPT-2 sizes small (117M), medium (345M), large (762M), and the full xl (1.5B) model, each with its own file set.
- •A working detector baseline baseline.py ships a reproducible detection model that OpenAI reported hitting accuracies in the mid-90s against Top-K 40 truncated generations.
- •Finetuning-evasion data a separate Amazon-review-finetuned GPT-2 sample set lets researchers test whether finetuning helps generated text evade detection.
- •Azure-hosted downloads download_dataset.py pulls every split from openaipublic.blob.core.windows.net, after OpenAI migrated the files off the original Google Cloud Storage bucket.
Detection benchmarks:
OpenAI's published baseline reaches accuracies in the mid-90s for Top-K 40 generations and mid-70s to high-80s (depending on model size) for unrestricted random sampling, with evidence that finetuning from a released model can help text evade detection.
How to use it:
Clone the repo, run download_dataset.py to pull the splits you need from Azure, then either train your own classifier on the labeled jsonl files or run baseline.py against detection.md's documented setup to reproduce OpenAI's numbers.
Limitations:
GitHub marks this repository archived, so there have been no commits, releases, or issue responses since December 2023: treat it as a frozen historical dataset, not an actively maintained tool. It only covers GPT-2-era text, so a detector trained purely on this data will not generalize to GPT-4, Claude, or Gemini-class generations. Data lives on Azure Blob Storage rather than in the repo itself, so downloads depend on that endpoint staying live, and OpenAI routes any data-removal requests for the underlying WebText corpus through a separate email address (webtextdata@openai.com) rather than a GitHub issue.
Our Verdict
This is OpenAI's original ground-truth dataset for detecting GPT-2-generated text, and it still functions as the standard reference researchers reach for when building or benchmarking machine-generated-text classifiers, even though GitHub has marked the repository archived since its last push in December 2023.
If you're an ML engineer or NLP researcher building a text-detection classifier, this dataset gives you a clean, pre-split, labeled 750K-document corpus per GPT-2 model size plus a reproducible mid-90s-accuracy baseline to benchmark against, all under the MIT license.
It is not the right choice if you need labeled data for modern LLM detection: the newest text here came out of a 2019-era 1.5B-parameter GPT-2, and today's detectors need to handle GPT-4, Claude, and Gemini-class writing that GPT-2 never touched. Skip if you're evaluating detectors against current-generation chatbots, want an actively maintained repo, or need data hosted directly in GitHub rather than pulled from Azure.
Frequently Asked Questions
What is the openai/gpt-2-output-dataset repo?
openai/gpt-2-output-dataset is OpenAI's official collection of GPT-2 model outputs, released alongside GPT-2's staged 2019 rollout to support research into detecting machine-generated text. It bundles 250,000 WebText human-written documents plus 250,000 random and 250,000 Top-K 40 generated samples for each of the four GPT-2 sizes (117M, 345M, 762M, and 1.5B parameters), split into train, validation, and test sets.
Is the GPT-2 output dataset free to use?
Yes. OpenAI released the dataset and its accompanying baseline.py detection code under the MIT license, so you can download, modify, and redistribute the data and code with no cost and minimal restriction. The repository has 2,026 GitHub stars and 546 forks as of 2026, reflecting continued research interest despite the project being archived.
How accurate is OpenAI's GPT-2 detector baseline?
OpenAI reports the baseline detector included in this repo reaches accuracies in the mid-90s for Top-K 40 truncated generations and mid-70s to high-80s, depending on model size, for unrestricted random generations. The team also found evidence that finetuning a released GPT-2 model can help its output evade detection, which is why a separate finetuned-sample set is included.
Is openai/gpt-2-output-dataset still maintained?
No, GitHub marks the repository as archived, with the last push recorded in December 2023 and no open issues being actively resolved. The dataset files themselves remain live and downloadable from Azure Blob Storage at openaipublic.blob.core.windows.net, so the data is still usable even though the code and documentation are frozen.
How does openai/gpt-2-output-dataset compare to newer AI-text-detection datasets?
openai/gpt-2-output-dataset only covers text generated by GPT-2, 2019-era models up to 1.5B parameters, so it is best for historical research or teaching detection fundamentals rather than benchmarking against modern chatbots. Choose this dataset when you specifically need GPT-2-era ground truth or OpenAI's original detector baseline; choose a newer corpus such as RAID or HC3 when you need labeled outputs from GPT-4, Claude, or Gemini-class models.
What is gpt-2-output-dataset?
openai/gpt-2-output-dataset is OpenAI's official dataset of GPT-2 model outputs, released under the MIT license to support machine-generated-text detection research. It's the archived, still-downloadable dataset researchers cite when benchmarking AI-text detectors against GPT-2-era writing.
How do I install gpt-2-output-dataset?
Visit the GitHub repository at https://github.com/openai/gpt-2-output-dataset for installation instructions.
What license does gpt-2-output-dataset use?
gpt-2-output-dataset uses the MIT license.
What are alternatives to gpt-2-output-dataset?
Explore related tools and alternatives on My AI Guide.
Open source & community-verified
MIT licensed: free to use in any project, no strings attached. 2,026 developers have starred this, meaning the community has reviewed and trusted it.
Reviewed by My AI Guide for relevance, quality, and active maintenance before listing.
