Multi-Domain AI/Human Text: Balanced Corpus for OOD and Adversarial Detection

A balanced binary detection corpus contains 45,000 text samples across nine named splits for evaluating AI-generated text classifiers. The dataset, created by author 'acmc', is structured to test in-distribution generalization and four independent out-of-distribution axes. It was last updated on Hugging Face in May 2026.

Use Cases

• Train binary classifiers to distinguish AI-generated from human-written text based on the balanced corpus.
• Benchmark model generalization on in-distribution (IID) and out-of-distribution (OOD) splits covering domain, language, and model shifts.
• Evaluate classifier robustness against adversarial attacks based on character, word, and paraphrase-level perturbations described in the splits.

Strengths

• The dataset provides nine distinct test splits, each containing 5,000 samples with a balanced 2,500 human and 2,500 AI texts.
• It is explicitly designed to evaluate four independent OOD axes: domain, language, model, and adversarial perturbations.
• The total corpus size across all splits is at least 45,000 text samples, based on the provided split details.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full dataset and training/validation splits is unknown, which may limit suitability assessment.
• The description metadata is limited; actual data quality and text sources require manual inspection after download.

Provenance

Source

huggingface

Freshness

Last updated 2026-05-31 16:13:04; freshness should be verified.