Download
The GPT-2 Output Dataset is a large collection of model-generated text, released by OpenAI alongside the GPT-2 research paper to study the behaviors and limitations of large language models. It contains 250,000 samples of GPT-2 outputs, generated with different sampling strategies such as top-k truncation, to highlight the diversity and quality of model completions. The dataset also includes corresponding human-written text for comparison, enabling researchers to explore methods for distinguishing machine-generated content from human-authored text. The repository provides scripts and metadata for working with the dataset, with the goal of supporting research in areas like detection, evaluation of text coherence, and analysis of generative models. While no active development is expected, the dataset remains a useful benchmark for tasks involving text classification, style analysis, and generative model evaluation.
Features
- 250,000 GPT-2 generated text samples across different prompts
- Includes both model outputs and human-written reference texts
- Generated using multiple sampling strategies (e.g., top-k truncation)
- Metadata and scripts provided for dataset exploration and processing
- Useful for studying detection of machine-generated vs human-written text
- Benchmark for evaluating generative models’ output quality and coherence