rounaksaha12/ai-det-test-human-refined

Clean Human Data for AI Detector benchmark

1,689 human-written web documents for evaluating AI-text detectors, drawn from FineWeb.

Read this before use

There is no text column, on purpose. The benchmark text is text_refined; text_original is the raw FineWeb extraction kept for provenance. A habitual ds["text"] raises ValueError: Column 'text' doesn't exist. rather than quietly returning the wrong one.

What are the discard rows? The file carries all 1,689 reviewed documents, including 39 the reviewer deemed questionable on the grounds of quality.

ds = load_dataset("rounaksaha12/ai-det-test-human-refined", split="test")

Provenance and cleaning

1. Sampled from 24 FineWeb shards across 8 pre-2021 CommonCrawl dumps (2013–2020), chosen as untouched by the training draw. Pre-2021 only, so the text predates widespread LLM generation.
2. Cleanliness judged by gemini-2.5-flash — only documents verdicted clean were kept.
3. Stratified at 50 documents per WebOrganizer topic, then boosted so thin document formats reach a usable count.
4. Scraping residue flagged by gemini-2.5-pro across 9 categories, with every cited span verified as a verbatim substring of the source.
5. Hand-reviewed: 789 documents were manually edited to strip residue; 39 further documents deemed questionable but retained here with discarded=True for auditability.

Columns

column	meaning
text_refined	the benchmark text — hand-refined where residue was found
text_original	as extracted by FineWeb, before refinement
discarded	True for documents the reviewer deemed to be of questionable quality
review_status	not_required / refined / discarded
was_refined	whether a human edited this document
topic_label / format_label	WebOrganizer classifiers (+ probabilities)
stratum	core (50/topic) or boost (format top-up)
flags_non_structural / flags_structural	residue categories the judge raised
flag_evidence	JSON: the verbatim span the judge cited per flag
judge1_* / judge2_model	cleanliness judge, residue judge
url, dump, date, warc_path	FineWeb provenance

Composition

All figures below cover all 1,689 rows, discarded ones included.

• Length (words, text_refined): median 486, mean 619, p90 1173, range 142–4793
• Topics: 23 — largest Finance & Business (119), smallest Fashion & Beauty (54)
• Formats: 22 — largest Personal Blog (318), smallest Structured Data (48)
• Crawls: 8, CC-MAIN-2013-20 to CC-MAIN-2020-24
• Strata: core 1,150, boost 539

Known limitations

• Document length varies ~2.5x by format (Kruskal–Wallis p ≈ 7e-13) but not by topic (p ≈ 0.14). Per-format comparisons are length-confounded; per-topic ones are not.
• FineWeb's PII scrubbing leaves placeholder tokens such as firstname.lastname@example.org in 61 documents. No LLM emits these, so they are a giveaway token if you pair this set with generated text.
• Whitespace is as FineWeb extracted it. Human FineWeb text never contains a blank line while LLM output usually does, so normalise whitespace across both classes or a detector will learn the formatting rather than the prose.
• boost rows over-represent thin formats by design; use stratum == 'core' for a distribution-faithful sample.

Usage

from datasets import load_dataset

ds = load_dataset("rounaksaha12/ai-det-test-human-refined", split="test")

texts = ds["text_refined"]      # note: text_refined, not text