jed-lee/mirm_style_datasets

MIRM Style Datasets

Datasets prepared for a CS8803 MIRM team project at Georgia Tech. The project trains one style embedding model for many style tasks, so each dataset here is shaped for contrastive training.

Each dataset is its own config. Every config shares the same four core columns, so the training code can treat them the same way. Some configs add metadata columns after those. More configs will be added over the coming weeks.

Usage

from datasets import load_dataset
train_data = load_dataset("jed-lee/mirm_style_datasets", "3lf", split="train")
val_data = load_dataset("jed-lee/mirm_style_datasets", "3lf", split="validation")

# raid is about 5 million rows, so streaming avoids a full download
raid_train = load_dataset("jed-lee/mirm_style_datasets", "raid", split="train", streaming=True)

Configs

• 3lf: formality (informal vs. formal), with train and validation

• raid: which model family wrote a text (or human), with train and validation

• argugpt: which GPT model wrote an essay (no human essays), with train, validation, test and ood

• outfox: human or which model wrote an essay, with train, validation and test

• raid_model, raid_ai and outfox_ai: the same rows as raid and outfox, labelled by exact model or by human vs AI

• valla_ccat50: which Reuters reporter wrote a news story, with train and validation

• valla_blogs: which blogger wrote a post, with train and validation

• wiki_auto: simple or standard English Wikipedia, with train and validation

• h4_stack_exchange: who wrote a Stack Exchange answer, with train and validation

Columns

These four are in every config and every split.

• text: one text, exactly as it appears in the source
• primary_id_int: the class to learn (its meaning depends on the config)
• secondary_category_int: shared content, such as a prompt or an aligned example. -1 means unknown and never counts as a shared group.
• source_record_id: traces the row back to the source

Integer ids are local to each config. In train they run from 0 with no gaps, which makes random access easy when building batches. Validation keeps the same labels. New groups in validation get ids after the training ones. The mappings for every id are in <config>/mappings.json.

3lf

The source pairs 1,500 example sentences with an informal and a formal rewrite of each. Each pair becomes two rows that share one secondary_category_int. That gives two texts with the same meaning in different registers, so a batch can learn formality without leaning on topic.

• primary_id_int: 0 = informal, 1 = formal
• secondary_category_int: the aligned pair
• source_record_id: 3lf-train-<position in train.json>-<informal|formal>

Split	Rows	Pairs	Informal	Formal
train	2,697	1,350	1,350	1,347
validation	300	150	150	150

No casual class. The 3LF paper describes three levels. The casual sentences come from GYAFC, which cannot be redistributed, so the public release leaves them out. They were not recreated here, so this config has two classes.

Preprocessing

The text is not changed. Spelling, capitalisation and punctuation stay as they were. Emojis and slang like "lol" stay too, because for style those details are the signal.

• Empty texts removed: 0
• Texts with both labels removed: 0
• Exact duplicates removed, first copy kept: 3
• Validation: 10% of pairs, seed 42, with each pair kept whole

Before saving, the script runs a set of checks. No text or pair may cross splits. Training ids must have no gaps, and every class needs at least two training texts. Every row must still match its source text and label. The script, mappings and counts are in the 3lf/ folder.

raid

RAID is a benchmark for detecting AI-generated text. It has human texts across eight domains, generations from 11 models written from the same prompts, and 11 adversarial attacks applied to both. Only the released training data is used. The official test split has no labels, and the extra split covers other domains.

• primary_id_int: the model family, see the table below
• secondary_category_int: the human source document (RAID source_id)
• source_record_id: RAID id

id	family	RAID models
0	human	human
1	cohere	cohere, cohere-chat
2	llama	llama-chat
3	mistral	mistral, mistral-chat
4	mpt	mpt, mpt-chat
5	openai	gpt3, chatgpt, gpt4
6	gpt2	gpt2

Base and chat versions of a model share a family. GPT-2 is the exception to grouping by company. It is a 2019 base model from an older generation, and it writes very differently from the later OpenAI models, so it has its own id. For plain human vs AI, use primary_id_int == 0 or the raid_ai config. The exact generator is still in the model column, and raid_model uses it as the label.

A group holds one human text, every generation made from its prompt, and every attacked copy of those. After the STEB drops, 66 train groups have no human text. So texts in one group share their prompt and differ in which family wrote them. That makes them good hard negatives.

Five metadata columns are kept as they are in RAID. They are not labels. model is the generator (human for human text) and domain is the genre. decoding and repetition_penalty are empty for human text. attack is none for unattacked text.

Split	Rows	Source docs	Human	Cohere	Llama	Mistral	MPT	OpenAI	GPT-2
train	4,475,959	12,034	131,353	516,246	525,765	1,029,010	995,792	770,732	507,061
validation	498,586	1,337	14,645	57,509	58,556	114,663	110,935	85,876	56,402

Still uneven, but much less. Human is the smallest family at about 1 to 8 against Mistral. As a human vs AI split it is 1 to 33, which is why raid uses the family. raid_ai has that split for training code that balances the classes.

Preprocessing

Only the whitespace at the very start and end of each text is removed. In the raw data 99.7% of GPT-3 texts start with a blank line and 99.7% of Cohere base texts start with a space, which no other model does. That is a label shortcut from how the texts were saved, not style. Everything inside the text stays as it is, including attacked text with homoglyphs, misspellings or zero-width characters.

• Empty or invisible texts removed: 473 (some rows were only zero-width spaces)
• Identical texts from two different families removed: 449
• Exact duplicates removed: 639,620. Most are attacks that changed nothing, such as the number attack on a text with no digits. The unattacked copy is kept.
• Kept on purpose: 3,382 texts under 50 visible characters. These include short model refusals like "Sorry, but I can't assist with that."
• 733 train rows are removed because STEB scores the same text, matched mostly in PAN25 collaborative. They are the human text and its attacked copies in 66 source groups, which keep their model texts but now have no human text. Validation keeps the 46 rows of 4 such groups, because no model trains on them. If you pick checkpoints on validation, those rows are STEB texts.
• Validation: 10% of source documents, seed 42. A whole group moves together, so every attacked copy stays with its original. Documents that share an identical text (301 of them) also go to the same split, so no text or its attacked copies can sit on both sides.

Before saving, the script checks that every attacked row points at an original with the same source_id. It also checks that no text, kept or removed, touches both splits and that training ids have no gaps. A separate check compared 4,000 random rows against the raw file for text, label and metadata.

raid_model, raid_ai and outfox_ai

These three configs are views of raid and outfox. Each holds the same rows, text, splits, groups and metadata as its parent, and only primary_id_int changes. STEB labels its AI tests by model, and a few only as human vs AI, so these views give the training code both. A view adds no new text, so don't count it as extra data.

• raid_model: which of RAID's 11 models wrote a text, or human. 0 = human, 1 = cohere, 2 = cohere-chat, 3 = llama-chat, 4 = mistral, 5 = mistral-chat, 6 = mpt, 7 = mpt-chat, 8 = gpt3, 9 = chatgpt, 10 = gpt4, 11 = gpt2. In train, human has 131,353 rows and each model has 246,436 to 525,765.
• raid_ai: 0 = human, 1 = AI. Human is about 3% of rows, so balance the classes when sampling.
• outfox_ai: 0 = human, 1 = AI (any of its three models).

Config	Split	Rows	Human	AI
raid_ai	train	4,475,959	131,353	4,344,606
raid_ai	validation	498,586	14,645	483,941
outfox_ai	train	53,247	10,047	43,200
outfox_ai	validation	1,842	342	1,500
outfox_ai	test	1,814	314	1,500

Each view's script copies its parent's files, changes only primary_id_int, and checks the new label against model and the row counts against the parent. A separate check on 30 Oct compared every column of every row with the parent.

argugpt

ArguGPT is a set of argumentative essays written by GPT models for real exam prompts (WECCL, TOEFL and GRE). The human essays are not released, because the authors do not have the copyright. So this config has no human class. The label is the model.

• primary_id_int: the model, see the table below. The 7 models in train get 0 to 6, and the 4 models that only appear in ood get 7 to 10.
• secondary_category_int: the essay prompt (ArguGPT prompt_id), so essays on the same prompt by different models share a group
• source_record_id: ArguGPT id

id	model
0	gpt-3.5-turbo
1	gpt2-xl
2	text-babbage-001
3	text-curie-001
4	text-davinci-001
5	text-davinci-002
6	text-davinci-003
7	bloomz-7b
8	claude-instant
9	flan-t5-11b
10	gpt4

Metadata columns: model, exam_type, temperature and score_level (empty in ood).

The official splits are kept, and dev is renamed to validation. ood is the authors' out-of-distribution set, with 4 models never seen in train plus gpt-3.5-turbo. The authors' validation and test splits mostly reuse training prompts: 294 of 344 validation essays and 296 of 346 test essays (86%) answer a prompt that is also in train. So they test new essays, not new topics. Only ood has no prompt overlap with train.

Split	Rows	Prompts
train	3,303	523
validation	344	82
test	346	82
ood	500	50

Model	train	validation	test	ood
gpt-3.5-turbo	639	87	86	100
gpt2-xl	249	52	55	0
text-babbage-001	366	60	60	0
text-curie-001	392	61	60	0
text-davinci-001	509	0	0	0
text-davinci-002	511	0	0	0
text-davinci-003	637	84	85	0
bloomz-7b	0	0	0	100
claude-instant	0	0	0	100
flan-t5-11b	0	0	0	100
gpt4	0	0	0	100

Preprocessing

GPT-2 was given one human-written sentence to continue from, so the first sentence of every gpt2-xl essay is human. That sentence is cut off here, otherwise human text would be labelled as GPT-2. 400 essays were cut. The 1 essay with no sentence end at all was dropped, and so were 44 essays left under 50 words after the cut, which is the authors' own minimum for GPT-2. After the cut, about 31% of GPT-2 essays open mid-thought ("However,", "This", "They"), against 16% or less for the other models. That is worth knowing when reading GPT-2 results. Apart from that, only edge whitespace is removed. There were no empty, conflicting or duplicate texts.

Some gpt-3.5-turbo essays still start with "As an AI language model" and many gpt4 essays start with "Title:". These are real model habits, so they stay.

outfox

OUTFOX pairs essays by US students in grades 6 to 12 with essays by three models, all answering the same problem statement.

• primary_id_int: 0 = human, 1 = chatgpt, 2 = flan_t5_xxl, 3 = text_davinci_003
• secondary_category_int: the problem statement, so the human essay and the three model essays for one prompt share a group
• source_record_id: outfox-<split>-<position in the file>-<model>

Model	train	validation	test
human	10,047	342	314
chatgpt	14,400	500	500
flan_t5_xxl	14,400	500	500
text_davinci_003	14,400	500	500

The model classes are the same size. Human is smaller after removing repeated essays, most of all in validation and test, because a repeat is kept in the earliest split. In train it is smaller again after the STEB drops below. The official splits are kept and valid is renamed to validation. The test-only attacked essays (OUTFOX attacker and DIPPER) are not included. 76 problem statements repeat across splits, because some prompts are generic.

Preprocessing

• Curly quotes become straight quotes, and non-breaking spaces become plain spaces. The human essays were saved with no curly quotes and many non-breaking spaces, and the model essays the other way round, so the file format alone gave the label away. No words change.
• Edge whitespace is removed.
• 2,774 repeated human essays are removed, keeping the earliest copy. Repeats are matched with whitespace collapsed, because the same student essay was often saved twice with only line breaks or spacing changed. Some of those copies sat in different splits.
• 1,923 human essays in train are removed because STEB scores the same essay: 1,895 in PAN25 collaborative, which reuses these essays, and 28 in GEDE. A model should not train on a text it is tested on. Validation and test keep theirs, because no model trains on them: every one of their 342 and 314 human essays is also in STEB, so dropping them would leave both splits with no human class. If you pick checkpoints on validation, those essays are STEB texts.
• Double spaces stay, since they are likely real typing habit.
• About 80% of chatgpt and text_davinci_003 essays stop mid-sentence, and about 30 to 40% drift into word salad near the end. The cause is not stated in the source. These essays are kept as they are, so keep this in mind when evaluating.

Every row was checked against the raw files for text, label, split and group.

valla_ccat50

CCAT50 is 50 Reuters reporters with 100 news stories each, from 1996 and 1997. It is one of the authorship sets in the VALLA benchmark. VALLA ships scripts but no data, so this config is rebuilt from the original UCI files. VALLA keeps the official test set and splits the rest by story. This config splits by reporter instead.

• primary_id_int: the reporter. Train reporters get 0 to 44 and validation reporters get 45 to 49.
• secondary_category_int: -1 for every row. The files come with no topic, so none is made up.
• source_record_id: the path in the UCI zip, <C50train|C50test>/<reporter>/<file>

Metadata columns: author (the reporter's name) and official_split (the original C50 train or test).

Split	Rows	Reporters	Stories per reporter
train	4,480	45	95 to 100
validation	500	5	100

Split by reporter. All of a reporter's stories go to one split, so validation tests reporters the model has never seen. The official C50 train and test halves hold the same 50 reporters, so that split is kept in official_split instead.

Topic follows the reporter. Each reporter covers a beat, such as Czech markets or Chinese politics, so topic alone can point to the author. RCV1 has topic codes for these stories, but almost every story has two or more, so they are not used here.

Preprocessing

• Windows line breaks become \n, and & becomes &.
• The desk sign-off at the end of 1,158 stories is removed, for example -- Prague Newsroom, 42-2-2423-0003 or ((--Aaron Pressman, 202-898-8312)). It is wire boilerplate with a phone number or email, so it gave the reporter away. When the last line is a currency note, the sign-off above it is removed and the note stays. Every distinct cut was checked by hand.
• Edge whitespace is removed.
• 20 repeated stories are removed, matched with whitespace collapsed. All were the same reporter filing the same story twice.

337 stories still end with a currency note like ($1=.6120 Pound), spread over 30 reporters. Every row was checked against the raw files for text, reporter and split.

valla_blogs

This is VALLA's Blogs50: the 50 bloggers with the most posts in the Blog Authorship Corpus (blogger.com, 2004). VALLA's script for it stops before saving, so this config rebuilds it. Posts under 100 characters are dropped and the top 50 bloggers by post count are kept. VALLA keeps the first copy of a post that two or more bloggers share, but here every copy is dropped. With the STEB skips below, these are not exactly VALLA's 50.

• primary_id_int: the blogger. Train bloggers get 0 to 43 and validation bloggers get 44 to 49.
• secondary_category_int: the blogger's self-reported industry, which VALLA calls topic. indUnk (unknown) is -1. It is set per blogger, not per post, so a blogger's own posts always share one group.
• source_record_id: <blogger id>-<position of the post in the blogger's file>

Metadata columns: author (blogger id), industry and date.

Split	Rows	Bloggers	Posts per blogger	Rows with industry -1
train	37,606	44	521 to 1,647	24,279
validation	4,251	6	549 to 931	2,066

id	industry	id	industry
0	Arts	5	Law
1	Education	6	Student
2	Fashion	7	Technology
3	Government	8	Communications-Media
4	Internet

Most industry groups hold only one blogger, and 31 of the 50 bloggers are indUnk. So the industry is a weak group, and most hard negatives will come from Student.

Split by blogger. All of a blogger's posts go to one split.

Bloggers kept apart from STEB. STEB tests on this same corpus, so 11 bloggers that would have made the top 50 are left out first. 5 are in STEB's blog_age and blog_gender test, which is every 20th file. 6 have a post that matches a STEB authbench_attribution_en test post (from MaoXun/AuthBench), meaning 50% or more of its 8-word runs. A second check with a different method found one match, a long run of "hahaha", and no real STEB test post.

Posts shared with STEB probing. STEB's probing_blog also uses posts from this corpus, and every one of the 50 bloggers here has posts in it. So leaving bloggers out can't avoid them. Instead, 2,854 posts are removed because STEB scores the same post: 2,852 in probing_blog and 2 that a PAN25 collaborative text copies. That's 2,570 train rows and 284 validation rows. They go after the split, so no blogger moves. Each blogger keeps 521 or more posts. Validation drops them too, since it has plenty of posts left.

Preprocessing

• The files are UTF-8 with some Windows-1252 bytes mixed in, mostly curly quotes. Those bytes are read as Windows-1252, so nothing is dropped or garbled.
• HTML escapes such as   and & are undone, and non-breaking spaces become plain spaces (same as outfox). NUL bytes from file damage are removed.
• Each line is stripped and blank lines are dropped, since the XML pads every post with them. urlLink stays, as that is how the corpus marks a link.
• Removed across the whole corpus:
  • 3,160 empty posts
  • 92,809 posts under 100 characters
  • 80,829 posts shared by two or more bloggers, such as quizzes, chain posts and mirrored blogs
  • 1,759 repeats within one blogger
• Removed after the split: the 2,854 posts STEB scores (see above)

One blogger ends 77% of their posts with the same quote. That is their own sign-off, so it stays. Every row was checked against the raw files for text, blogger, industry and date.

wiki_auto

Wiki-Auto pairs English Wikipedia articles with their Simple English Wikipedia versions, from the September 2019 dump. Its authors aligned the sentences in each pair with their own model. Each aligned sentence becomes one row, and the sentences of one article pair share a secondary_category_int. So a standard sentence and its simpler version sit in the same group, and a batch can learn simplicity without leaning on topic.

• primary_id_int: 0 = standard English Wikipedia, 1 = Simple English Wikipedia
• secondary_category_int: the article pair
• source_record_id: the Wiki-Auto sentence id, <normal|simple>-<article id>-<paragraph>-<sentence>

Metadata columns: pair_id (<part1|part2>-<Wiki-Auto example id>) and title (the English article title).

Split	Rows	Article pairs	Standard	Simple
train	925,855	107,047	423,522	502,333
validation	102,904	11,949	46,924	55,980

The label is the edition. A sentence is labelled by the Wikipedia it comes from, so the label is only a rough simplicity signal. Simple sentences average 15 words and standard ones 22. Some aligned pairs differ by a fact or a name, not by simplicity.

Only aligned sentences. Sentences with no alignment are left out. That keeps both classes about the same content and about the same size. A sentence aligned to two or more others is one row.

Same sentence on both sides. A sentence that appears in both editions is removed, every copy. Sentences count as the same when their letters and digits match, ignoring case, spacing, punctuation and accents. So a sentence that only changed ( , ; to (, ; or a straight apostrophe to a curly one is removed too. That kind of change is not simpler writing, and it would let the format give the label away.

Kept apart from STEB. Seven STEB tests have Wikipedia text. Five are ASSET (validation and test), STEL, M4 Wikipedia, MAGE squad and the first 1,000 Pile Wikipedia articles. The other two are the CORE and X-GENRE web genre sets, which include pages copied from Wikipedia. Any article pair with a sentence in those is removed, 1,473 pairs in all. A sentence matches if it has the same words as a STEB text, or 90% or more of its 8-word runs are in the STEB texts. It also matches if it is a small edit of one STEB sentence: 40 or more letters, 90% the same, with the same names and numbers. The names and numbers rule stops template lines like "As of the 2010 United States Census, the population was 4,500." from matching every town. A little overlap remains. A STEB sentence with a respelled name, like "Shostakovitch" for "Shostakovich", is not caught. A second check with a different method found 2 simple sentences just under the 90% line, such as "Some not sure sources say that he was also in Paris between 1308 and 1310." against "Some speculative sources say…" in ASSET. Short stock lines such as "The opera was well received." also appear in STEB texts.

Drop list. The AI-detection experiment's drop list checks every STEB test, not just these seven. It finds 170 more sentences across 18 tests. 70 sit inside PAN25 collaborative texts, and 67 are in the seven tests above but slipped past the first check. Their 170 article pairs are removed too, 3,293 sentences. That happens after the split, so no pair moves. Validation drops them as well.

Preprocessing

The text is not changed. It keeps a few gaps that the Wikipedia text extractor left where it dropped some text, for example Polikarpov ( ; 9 June O.S. 28 May 1892, where text inside the brackets was dropped. Both editions have them, but 4.2% of standard rows have such a gap, against 0.8% of simple rows. A few rows are sentence-splitting mistakes that end on a comma.

• Empty texts removed: 0
• Article pairs with a STEB sentence removed: 1,473 (48,527 sentences)
• Article pairs with a sentence on the drop list removed, after the split: 170 (3,293 sentences)
• Sentences on both sides removed, every copy: 181,838
• Repeats on one side removed, first copy kept: 9,502. Most are template lines like "It is found in the Yonne department in the center of France."
• Validation: 10% of sentences, by article pair, seed 42

Every row was checked against the raw files for text, label, article pair and title.

h4_stack_exchange

Answers from 171 Stack Exchange sites, from the Stack Exchange data dump as copied in HuggingFaceH4/stack-exchange-preferences. Each row is one answer, and the label is who wrote it. Stack Overflow and the meta sites are left out. Stack Overflow alone holds 24M of the 32M answers, and most of them are code.

• primary_id_int: the author. User ids are local to a site, so one person on two sites has two ids.
• secondary_category_int: the question the answer replies to. Answers to one question share a topic but not an author, so they make good hard negatives.
• source_record_id: <site>/<answer id>

Metadata columns: site, author (the display name), author_profile, answer_url and question_url. They are there for the CC BY-SA attribution (see Licenses). mappings.json maps each author id to its site and Stack Exchange user id, and each question id to its site and question id.

Split	Rows	Authors	Answers per author	Questions
train	5,265,569	350,894	2 to 12,313, median 3	2,209,429
validation	584,940	39,610	2 to 5,336, median 3	505,665

Split by author. All of an author's answers go to one split. So one question's answers can sit in both splits, and 460,332 questions do. 90% of validation rows answer a question that also has answers in train. In validation only 24% of rows share their question with another validation row, so the question is a weak group there.

Some authors write a lot. There is no cap per author. The top 1% of authors wrote 39% of train rows, and one wrote 12,313 answers. Math Stack Exchange is 17% of all rows.

Kept apart from STEB. Seven STEB tests are built from Stack Exchange: stackexchange_retrieval, probing_stackexchange, stackexchange_politeness, AuthBench's Stack Exchange texts, the first 1,000 Pile Stack Exchange documents, and PAN18 and PAN22 style change. An answer that holds one of their texts is removed. A text of 8 or more words matches when 90% or more of its 8-word runs are in the answer and a second check finds 90% or more of its words in order. Long documents are checked sentence by sentence. A shorter text from retrieval, probing, politeness or AuthBench must equal the whole answer. STEB's copies come from the markdown source, with links, quotes and code swapped for placeholders. So runs that cross one of those don't count, unless no run would be left. That removes 16,388 answers, most of them matching PAN22 (7,187), PAN18 (4,016), probing (2,803) or retrieval (2,027).

stackexchange_retrieval and AuthBench test authors, so they also remove authors. Retrieval's user labels are not the sites' user ids. But each STEB user's matched texts mostly land on one author. So an author goes when they hold 2 or more of one STEB user's matched texts, or all of them. AuthBench removes the author of any match, and none of its texts match. That takes 489 authors and 9,568 more answers.

The other STEB tests that might hold Stack Exchange text are CORE, X-GENRE and PAN23 to PAN26. They mostly share quotes, Bible verses and stock lines with answers. So there an answer goes only when 2 or more sentences of one STEB text are in it and that text holds half or more of the answer. An example is a web page that copied an answer. That's 48 answers. In every test, a text found under 3 or more authors never counts. It's a stock phrase like "If it ain't broke, don't fix it." The Pile's other parts, such as its web pages, are not checked.

Preprocessing

The answers are HTML. They become plain text, close to what a browser shows.

• Quotes (<blockquote>) are dropped, since they are someone else's words. Spoilers look like quotes in the HTML but are the author's own, so they stay.
• Code blocks keep their line breaks and indentation. Maths stays as its LaTeX source, like $x^2$.
• Images and embeds are dropped. Links become their text. Struck-through text stays.
• Each list item and table row gets its own line. Table cells are joined by a space.
• Outside code, a line break in the HTML is layout, so it becomes one space. A dropped image leaves one space too, or nothing when it sat inside a word. Old and new Stack Exchange pages lay out the same answer differently. Without this, the layout would give away the year and so the author. Spaces the author typed, like two after a full stop, stay.
• Blank lines are capped at one outside code. Outside code, non-breaking spaces become plain spaces. Spaces at the end of a line go everywhere, and so does the indent of an answer's very first line.

Removed, in this order:

• Copies: 1,080,445. H4 repeats some answers under 2 to 19 questions. Each answer is kept once, under its lowest question id. A check of 100 such answers on the Stack Exchange API found that is the real one every time.
• Answers by "Community" (user -1): 207,902
• Empty after conversion, such as quote-only answers: 1,731
• STEB answers: 16,388, then 9,568 by the 489 STEB authors
• Same text under two authors, every copy: 5,108. Texts count as the same when their letters and digits match (as in wiki_auto).
• Repeats by one author, first kept: 3,077
• Authors left with one answer: 939,314 answers. An author needs two to form a pair.
• Validation: 10% of answers, by author, seed 42

There is no minimum length, and 0.9% of answers are under 50 characters. 525 answers keep control characters, mostly code golf. For 2,796 authors with 3 or more answers, half or more of their answers end with the same line. That's 59,598 rows (1.0%). Some lines are site-wide, like \end{document} on TeX or "Try it online!" on Code Golf. Others are habits, like "Hope this helps." They stay.

A separate check re-read every kept answer from the raw files with a second HTML parser. Ignoring whitespace, all but 714 (0.01%) match. Those are broken HTML, such as badly nested quotes, where the two parsers recover differently. The check also gave every removed answer a reason, and the rows left matched the build exactly. 200 random answers were also checked on the live sites. All are still up under the same author and question, and 199 match exactly.

Dataset Description

3lf. Hyojeong Yu, Hyukhun Koh, Minsung Kim and Kyomin Jung. Casual as an Anchor: Resolving Supervision Misalignment in Formality Transfer Dataset. arXiv:2605.29365, 2026.

Data obtained from stellahj/3lf, which is built from GYAFC examples. Thank you to the 3LF authors for making their rewrites public.

raid. Liam Dugan, Alyssa Hwang, Filip Trhlík, Andrew Zhu, Josh Magnus Ludan, Hainiu Xu, Daphne Ippolito and Chris Callison-Burch. RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors. ACL 2024, pages 12463 to 12492.

Data obtained from liamdugan/raid.

argugpt. Yikang Liu, Ziyin Zhang, Wanyang Zhang, Shisen Yue, Xiaojing Zhao, Xinyuan Cheng, Yiwen Zhang and Hai Hu. ArguGPT: evaluating, understanding and identifying argumentative essays generated by GPT models. arXiv:2304.07666, 2023.

Data obtained from huhailinguist/ArguGPT, commit cfbd00e.

outfox. Ryuto Koike, Masahiro Kaneko and Naoaki Okazaki. OUTFOX: LLM-Generated Essay Detection Through In-Context Learning with Adversarially Generated Examples. AAAI 2024, volume 38, issue 19, pages 21258 to 21266.

Data obtained from ryuryukke/OUTFOX, commit 8dd6bfd. The human essays come from the Kaggle Feedback Prize.

valla_ccat50 and valla_blogs. Jacob Tyo, Bhuwan Dhingra and Zachary C. Lipton. On the State of the Art in Authorship Attribution and Authorship Verification. arXiv:2209.06869, 2022.

CCAT50 data obtained from the UCI Machine Learning Repository, Reuter_50_50 (Zhi Liu, 2006). Blog data obtained from barilan/blog_authorship_corpus, from Jonathan Schler, Moshe Koppel, Shlomo Argamon and James Pennebaker, Effects of Age and Gender on Blogging, AAAI Spring Symposium 2006.

wiki_auto. Chao Jiang, Mounica Maddela, Wuwei Lan, Yang Zhong and Wei Xu. Neural CRF Model for Sentence Alignment in Text Simplification. ACL 2020, pages 7943 to 7960.

Data obtained from chaojiang06/wiki-auto, the full article-pair files (wiki-auto-all-data) from the authors' Dropbox, pinned by SHA-256.

h4_stack_exchange. Nathan Lambert, Lewis Tunstall, Nazneen Rajani and Tristan Thrush. HuggingFace H4 Stack Exchange Preference Dataset. 2023.

Data obtained from HuggingFaceH4/stack-exchange-preferences, revision c7bda74. The answers come from the Stack Exchange data dump and were written by Stack Exchange users.

Licenses

• 3lf: the source's terms, see stellahj/3lf
• raid: MIT, see liamdugan/raid
• argugpt: the GitHub repo has no license file, and the authors' HF copy lists only "cc". Treat it as research use and check with the authors before any other use.
• outfox: Apache-2.0, see ryuryukke/OUTFOX. The human essays come from the Kaggle Feedback Prize, which has its own terms.
• valla_ccat50: CC BY 4.0, see Reuter_50_50
• valla_blogs: the corpus may be freely used for non-commercial research purposes, see barilan/blog_authorship_corpus
• wiki_auto: Wiki-Auto has no license of its own. The Wikipedia text is CC BY-SA 3.0.
• h4_stack_exchange: CC BY-SA 4.0. The content is from the Stack Exchange Network. Each row names its author and links to their profile and to the question, as the license asks.

Citation

Please cite the sources for the configs you use: the 3LF paper and GYAFC for 3lf, RAID for raid, ArguGPT for argugpt, OUTFOX for outfox, VALLA plus the source corpus for valla_ccat50 and valla_blogs, Wiki-Auto for wiki_auto, and H4 Stack Exchange for h4_stack_exchange.

@misc{yu2026casualanchorresolvingsupervision,
  title         = {Casual as an Anchor: Resolving Supervision Misalignment in Formality Transfer Dataset},
  author        = {Hyojeong Yu and Hyukhun Koh and Minsung Kim and Kyomin Jung},
  year          = {2026},
  eprint        = {2605.29365},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2605.29365}
}

@inproceedings{rao-tetreault-2018-dear,
  title     = {Dear Sir or Madam, May {I} Introduce the {GYAFC} Dataset: Corpus, Benchmarks and Metrics for Formality Style Transfer},
  author    = {Rao, Sudha and Tetreault, Joel},
  booktitle = {Proceedings of the 2018 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers)},
  year      = {2018},
  pages     = {129--140},
  url       = {https://aclanthology.org/N18-1012}
}

@inproceedings{dugan-etal-2024-raid,
  title     = {{RAID}: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors},
  author    = {Dugan, Liam and Hwang, Alyssa and Trhl{\'\i}k, Filip and Zhu, Andrew and Ludan, Josh Magnus and Xu, Hainiu and Ippolito, Daphne and Callison-Burch, Chris},
  booktitle = {Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year      = {2024},
  pages     = {12463--12492},
  url       = {https://aclanthology.org/2024.acl-long.674}
}

@misc{liu2023argugpt,
  title         = {ArguGPT: evaluating, understanding and identifying argumentative essays generated by GPT models},
  author        = {Yikang Liu and Ziyin Zhang and Wanyang Zhang and Shisen Yue and Xiaojing Zhao and Xinyuan Cheng and Yiwen Zhang and Hai Hu},
  year          = {2023},
  eprint        = {2304.07666},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL}
}

@inproceedings{koike2024outfox,
  title     = {OUTFOX: LLM-Generated Essay Detection Through In-Context Learning with Adversarially Generated Examples},
  author    = {Koike, Ryuto and Kaneko, Masahiro and Okazaki, Naoaki},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  volume    = {38},
  number    = {19},
  pages     = {21258--21266},
  year      = {2024},
  doi       = {10.1609/aaai.v38i19.30120}
}

@misc{tyo2022valla,
  title         = {On the State of the Art in Authorship Attribution and Authorship Verification},
  author        = {Jacob Tyo and Bhuwan Dhingra and Zachary C. Lipton},
  year          = {2022},
  eprint        = {2209.06869},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL}
}

@misc{liu2006reuter5050,
  title     = {{Reuter\_50\_50}},
  author    = {Liu, Zhi},
  year      = {2006},
  howpublished = {UCI Machine Learning Repository},
  doi       = {10.24432/C5DS42}
}

@inproceedings{schler2006effects,
  title     = {Effects of Age and Gender on Blogging},
  author    = {Schler, Jonathan and Koppel, Moshe and Argamon, Shlomo and Pennebaker, James W},
  booktitle = {AAAI Spring Symposium: Computational Approaches to Analyzing Weblogs},
  volume    = {6},
  pages     = {199--205},
  year      = {2006}
}

@inproceedings{jiang-etal-2020-neural,
  title     = {Neural {CRF} Model for Sentence Alignment in Text Simplification},
  author    = {Jiang, Chao and Maddela, Mounica and Lan, Wuwei and Zhong, Yang and Xu, Wei},
  booktitle = {Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
  year      = {2020},
  pages     = {7943--7960},
  doi       = {10.18653/v1/2020.acl-main.709}
}

@online{h4stackexchange,
  author = {Lambert, Nathan and Tunstall, Lewis and Rajani, Nazneen and Thrush, Tristan},
  title  = {HuggingFace H4 Stack Exchange Preference Dataset},
  year   = 2023,
  url    = {https://huggingface.co/datasets/HuggingFaceH4/stack-exchange-preferences}
}