Datasets:
Request access to the V5 evaluation set
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access is granted case by case. By requesting access you agree to the Winston AI Evaluation Dataset License: research and evaluation use only, no commercial use, no redistribution, no training on this data.
Log in or Sign Up to review the conditions and access this dataset content.
V5 English AI vs Human Evaluation Set (12k)
English test set used for the v5 document-level AI-vs-human classification evaluation.
12,000 documents: 6,000 human (label = 0) and 6,000 AI-generated (label = 1).
Split: test.
Do not train on this set. It exists to evaluate detectors; any model that has seen it will report inflated numbers.
What's in a row
| Field | Type | Notes |
|---|---|---|
id |
int | Unique row id |
text |
string | Document text |
label |
class | human (0) / ai (1) |
model |
string | LLM that generated the text for AI rows; N/A for human rows |
category |
string | Domain (see below) |
Sampling
- 6,000 human rows, balanced by category: 600 each of
books,college-essays,medical,news,poems,political,recipes,reviews,social-media,wikipedia. - 6,000 AI rows, balanced by generator (20 generators; counts are not perfectly even where a generator's pool was small).
- Minimum 500 characters, English only.
Human sources
Europarl, Project Gutenberg, Wikipedia, RealNews, Signal Media, PubMed abstracts, Reddit, Food.com, IMDb, IvyPanda, College Essay Advisors, Poetry Foundation.
AI generators
Claude Haiku 4.5, Sonnet 4.6, Sonnet 5 · DeepSeek V4 Flash / Pro · Gemini 3.1 Pro, 3.5 Flash, 3.5 Flash-Lite, 3.6 Flash · GPT-4o, 5.4, 5.6 Luna, 5.6 Terra · Grok 4.1 Fast, Grok 4.5 · Mistral Nemo · Qwen 3.7 Flash / Plus, Qwen 3.8 27B / Max.
Categories
abstracts, books, coding, college-essays, corpus, email, fiction, legal,
medical, news, poems, political, recipes, report, reviews, scientific,
social-media, wikipedia.
AI rows skew toward corpus (general web text); human rows are flat at 600 per category.
Intended use
Benchmarking document-level AI-text detectors. Classification is binary on label; the
category and model columns are there to break results down by domain and generator.
License
Access is granted on request and governed by the Winston AI Evaluation Dataset License: research and evaluation use only, no commercial use, no redistribution, no training. Underlying human-written texts remain subject to the licenses of their original sources.
- Downloads last month
- 11