ITAIS — Is This AI Slop? (v2)

Human vs AI-generated text corpus for training the ITAIS pile-resemblance detector. Doc-level binary labels are provenance-based: a row is label=1 iff a machine was watched generating it (or a public corpus shipped it as machine-generated). No model-judged labels decide ground truth.

Files

file	rows	what
train_all.parquet	786k	v1 corpus (legacy)
v2_train_labeled.parquet	222,067	v2 train — see below
v2_holdout_labeled.parquet	6,547	held-out eval, split-safe
v2_holdout_paraphrase_labeled.parquet	~1k	paraphrase stress test
v2_holdout_mixed.parquet	small	mixed-domain probe
v2_holdout_unseen_model_labeled.parquet	~300	generator-disjoint eval

v2 train composition

• 115,571 human (label=0): writingprompts (62.0k), gutenberg books (14.6k), wiki intros (14.5k), blogs (7.5k), plus human halves of hc3/beemo/raid pairs
• 106,496 AI (label=1) across 27 registers:
  • public corpora: m4_* (38.7k), wiki_intro_gpt, hc3_*_gpt, beemo_ai, raid
  • coai (30k; source models unnamed → generator=unknown)
  • contrastive generations (12,803): rewrite_pair (9,992) + respond_pair (2,811) written by gemma-4-26b, laguna-s-2.1 and ox-alpha-free (see generator column)

Schema

text, label (0=human, 1=AI), register, spans, source_dataset, generator, generation_method (direct/rewrite/respond), decoding, split_hint.

split_hint is a pair-key (para:<id> / premise:<id> / fictpair:<id>) joining each generated story to its human source text — keep both sides in the same split to avoid leakage.

Known caveats

• coai labels were verified by LLM rubric + audit; generators unnamed upstream
• hc3 length-shortcut risk mitigated by dropping over-long AI answers (worst family ratio ≤1.5)
• spans column reserved (ontology pattern hits); currently empty — deterministic matcher validated as non-discriminative and excluded from labeling

Intended use

Train/eval doc-level AI-text detectors. Not for determining authorship of real-world text.