As large language models converge on similar stylistic patterns, automatically distinguishing "stereotypical" AI-style prose requires human judgments about perceived sloppiness rather than binary provenance labels. This dataset preserves those human perceptions by collecting slider-based votes from contributors and keeping per-vote records so disagreement and nuance remain available for modelling and analysis.
What Sets It Apart
- Continuous human judgments: contributors rate samples on a -1/0/+1 slider, and each vote is stored rather than collapsed to a single label, enabling models to learn uncertainty and annotator disagreement.
- Provenance metadata: each annotation includes source_dataset, source_row_id and content_hash so researchers can trace examples back to public sources or filter by origin.
- Lightweight and extensible: packaged as a small JSON/Tabular dataset on Hugging Face with an MIT license and a continuously growing collection via the SlopFinder interface.
- Focused signal: labels target perceived stylistic markers of "AI slop" (e.g., purple prose, clichéd phrasing) rather than factual correctness or verified AI provenance.
Who It's For and Trade-offs
Great fit if you want to train or evaluate classifiers that predict human-perceived AI-like writing, study annotator disagreement about style, or probe stylistic features correlated with perceived "slop." Look elsewhere if you need objective ground-truth of model provenance (this dataset captures subjective perception), large-scale balanced corpora for production classifiers, or datasets annotated for factuality/stance rather than stylistic perception.
Where It Fits
This dataset complements automated slop-detection tools and other curated "slop" collections: it is smallest and most human-centric in preserving per-vote information, making it useful for research into label noise, annotator calibration, and models that predict consensus/confidence rather than binary outputs.
