AI-Generated Text Detection Benchmark Across 11 LLMs and Genres

RAID is a benchmark dataset for evaluating machine-generated text detectors, containing over 10 million documents. It spans 11 large language models, 11 genres, 4 decoding strategies, and 12 adversarial attacks. The dataset was created by liamdugan and updated in September 2024.

Use Cases

• Benchmark detector robustness across 12 adversarial attacks documented in the dataset.
• Evaluate detection performance variance across 11 distinct text genres.
• Analyze the impact of 4 different decoding strategies on detector accuracy.
• Compare generated text characteristics from 11 different large language models.

Strengths

• Contains over 10 million documents, providing a large-scale evaluation corpus.
• Covers 11 distinct text genres for domain-specific analysis.
• Includes text generated by 11 different large language models for model comparison.
• Incorporates 12 adversarial attacks to test detector robustness.

Limitations

• Dataset composition and potential class imbalances across 11 LLMs and 11 genres are not detailed.
• The raw description lacks specifics on document length, column structure, or data format.
• As a benchmark, it may be optimized for evaluation rather than general model training.

Provenance

Source

liamdugan on Hugging Face

Collection Method

Constructed as a shared benchmark for detector evaluation; specific collection methodology not detailed.

Freshness

Last updated in September 2024.