Hugging Face's logo

Datasets:
CoderBak
/
anagram-corpus-nc

Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
multilingual
Size:
1M - 10M
ArXiv:
Tags:
ai-text-detection
human-written-text
prompts
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
anagram-corpus-nc / README.md
CoderBak's picture
CoderBak
Dataset card, licences, sources
d1b2bb4 verified
preview code
|
Raw History Blame Contribute Delete
67.3 kB
metadata 
pretty_name: Anagram Corpus (non-commercial part)
license: other
license_name: anagram-corpus-nc-mixed
license_link: LICENSE.md
language:
  - en
  - multilingual
tags:
  - ai-text-detection
  - human-written-text
  - prompts
configs:
  - config_name: acl_ocl
    data_files:
      - split: full
        path: data/acl_ocl/full-*.parquet
  - config_name: aeslc
    data_files:
      - split: train
        path: data/aeslc/train-*.parquet
      - split: dev
        path: data/aeslc/dev-*.parquet
      - split: test
        path: data/aeslc/test-*.parquet
  - config_name: corefl
    data_files:
      - split: learners
        path: data/corefl/learners-*.parquet
      - split: native
        path: data/corefl/native-*.parquet
  - config_name: deceptive_opinion_spam
    data_files:
      - split: full
        path: data/deceptive_opinion_spam/full-*.parquet
  - config_name: echr
    data_files:
      - split: train
        path: data/echr/train-*.parquet
      - split: dev
        path: data/echr/dev-*.parquet
      - split: test
        path: data/echr/test-*.parquet
  - config_name: ellipse
    data_files:
      - split: train
        path: data/ellipse/train-*.parquet
      - split: test
        path: data/ellipse/test-*.parquet
  - config_name: ellipse_excluded
    data_files:
      - split: full
        path: data/ellipse_excluded/full-*.parquet
  - config_name: hupd
    data_files:
      - split: full
        path: data/hupd/full-*.parquet
  - config_name: ice
    data_files:
      - split: full
        path: data/ice/full-*.parquet
  - config_name: icle_locness
    data_files:
      - split: full
        path: data/icle_locness/full-*.parquet
  - config_name: idealens
    data_files:
      - split: human_rewrite
        path: data/idealens/human_rewrite-*.parquet
  - config_name: keystroke_learner_writing
    data_files:
      - split: full
        path: data/keystroke_learner_writing/full-*.parquet
  - config_name: keystroke_learner_writing_keystrokes
    data_files:
      - split: full
        path: data/keystroke_learner_writing_keystrokes/full-*.parquet
  - config_name: l1_signals
    data_files:
      - split: acl_postllm
        path: data/l1_signals/acl_postllm-*.parquet
      - split: acl_prellm
        path: data/l1_signals/acl_prellm-*.parquet
      - split: acl_prenn
        path: data/l1_signals/acl_prenn-*.parquet
  - config_name: methods
    data_files:
      - split: full
        path: data/methods/full-*.parquet
  - config_name: moviesum
    data_files:
      - split: train
        path: data/moviesum/train-*.parquet
      - split: test
        path: data/moviesum/test-*.parquet
      - split: validation
        path: data/moviesum/validation-*.parquet
  - config_name: no_robots
    data_files:
      - split: train
        path: data/no_robots/train-*.parquet
      - split: test
        path: data/no_robots/test-*.parquet
  - config_name: pelic
    data_files:
      - split: full
        path: data/pelic/full-*.parquet
  - config_name: persuade_2
    data_files:
      - split: train
        path: data/persuade_2/train-*.parquet
      - split: test
        path: data/persuade_2/test-*.parquet
  - config_name: sharechat
    data_files:
      - split: train
        path: data/sharechat/train-*.parquet

Anagram Corpus — non-commercial part

Human-written text, real prompts people wrote, and the generation methods of published papers, gathered from public datasets into one format. It is the input of anagram-bench, a benchmark for AI-text detectors: the benchmark samples from here, asks a language model to write or edit text, and labels the results.

Everything in this repository is for non-commercial use only (CC BY-NC, BY-NC-SA, BY-NC-ND, or quoted material). The commercially usable part is CoderBak/anagram-corpus; load both for the full corpus. The two never hold the same rows.

This version holds 4,854,732 rows in 20 subsets (one per source part). It is being built source by source; more subsets will follow.

Loading 

from datasets import load_dataset
ds = load_dataset("CoderBak/anagram-corpus-nc", "acl_ocl")   # subset = source; splits = the source's own splits

Each subset keeps its source's own splits (train, test, ...). A source with no official split has one split, full. Every row records its source's licence in license.

Subsets

Subset Record type Rows Splits Licence Source
acl_ocl documents 71,021 full 71,021 CC BY-NC-SA 3.0, CC BY-NC-SA 4.0 ACL OCL: The ACL Anthology Open Corpus
aeslc documents 18,302 dev 1,960, test 1,906, train 14,436 CC BY-NC-SA 4.0 AESLC: Annotated Enron Subject Line Corpus
corefl documents 5,638 learners 3,940, native 1,698 CC BY-NC-ND 3.0 COREFL: Corpus of English as a Foreign Language
deceptive_opinion_spam documents 1,600 full 1,600 CC BY-NC-SA 3.0 Deceptive Opinion Spam Corpus v1.4
echr documents 11,478 dev 1,380, test 2,998, train 7,100 ECtHR terms: reproduction free of charge with the source acknowledged, for private, information or education use; commercial use needs written permission ECHR: Neural Legal Judgment Prediction in English (European Court of Human Rights cases)
ellipse documents 6,482 test 2,571, train 3,911 CC BY-NC-SA 4.0 ELLIPSE: English Language Learner Insight, Proficiency and Skills Evaluation Corpus
ellipse_excluded documents 2,400 full 2,400 CC BY-NC-SA 4.0 ELLIPSE: English Language Learner Insight, Proficiency and Skills Evaluation Corpus
hupd documents 4,518,254 full 4,518,254 CC BY-NC-SA 4.0 HUPD: The Harvard USPTO Patent Dataset
ice documents 902 full 902 CC BY-NC-SA 3.0 ICE-Nigeria: the Nigerian component of the International Corpus of English
icle_locness documents 500 full 500 CC BY-NC 4.0 ICLE500: a 500-essay open sample of the International Corpus of Learner English (ICLEv3)
idealens documents 50 human_rewrite 50 CC BY-NC-SA 4.0 TwiceTold (IdeaLens): human stories written from AI outlines
keystroke_learner_writing documents 1,006 full 1,006 CC BY-NC-SA 4.0 KUPA-KEYS: Logging Keystrokes in Writing by English Learners
keystroke_learner_writing_keystrokes logs 2,012 full 2,012 CC BY-NC-SA 4.0 KUPA-KEYS: Logging Keystrokes in Writing by English Learners
l1_signals documents 443 acl_postllm 31, acl_prellm 12, acl_prenn 400 CC BY-NC-SA 3.0, Quoted from the paper (no licence; non-commercial research use with attribution), CC BY-NC-ND 4.0 Can We Still Trace L1 Signals? (academic abstracts labelled with the authors' native language)
methods methods 676 full 676 Quoted from the paper (no licence; non-commercial research use with attribution), CC BY-NC-SA 4.0, CC BY-NC-ND 4.0, CC BY-NC 4.0 anagram-bench generation methods
moviesum documents 2,200 test 200, train 1,800, validation 200 CC BY-NC 4.0 MovieSum: An Abstractive Summarization Dataset for Movie Screenplays
no_robots prompts 10,000 test 500, train 9,500 CC BY-NC 4.0 No Robots
pelic documents 46,204 full 46,204 CC BY-NC-ND 4.0 PELIC: University of Pittsburgh English Language Institute Corpus
persuade_2 documents 25,996 test 10,402, train 15,594 CC BY-NC-SA 4.0 PERSUADE 2.0
sharechat prompts 129,568 train 129,568 CC BY-NC 4.0 ShareChat: A Dataset of Chatbot Conversations in the Wild

Columns

Every row has: id (/); source (source name (= subset)); split (the source's own split, or full when it has none); license (licence of this row); source_record_id (the record's id in the source release); source_url (link to the original record, if any); meta (the source's other fields, as released (JSON)).

documents: text (the human-written text (normalised: see below)); title (title, if the source has one); prompt (the assignment or question the writer answered, if the source gives one); genre (register of the text, as labelled for this corpus); date (date written or published (ISO; as precise as the source allows)); pre_chatgpt (date before 2022-11-30 (null if unknown or ambiguous)); author_id (author id as released by the source); parent_id (thread / document the text belongs to); language (fastText lid.176 language code of the text); language_score (its probability); n_words (whitespace words); text_hash (sha1 of the lower-cased, whitespace-collapsed text).

logs: parent_id (id of the row this record belongs to); format (format of the record); data (the record exactly as released).

methods: kind (template | instruction | attack | classifier); task (the anagram-bench task that uses it); name (template, list or attack id); text (the template or instruction, exactly as quoted); system (system message of a template); placeholders (slots a template expects); params (settings of an attack (JSON)); verbatim (true if copied word for word from the source); paper (source paper); paper_url (its URL); location (where in the paper or repository).

prompts: prompt (the first user turn); system (system prompt, if any); messages (the conversation as {role, content}; assistant content is null when its licence differs from the row's); n_turns (number of user turns); human_response (a human-written answer to the prompt, if the source has one); category (task category given by the source); model (chatbot that answered (chat logs)); date (date of the conversation (ISO)); pre_chatgpt (date before 2022-11-30); language (fastText lid.176 language code of the first turn); language_score (its probability); n_words (words in the first turn); text_hash (sha1 of the first turn, lower-cased, whitespace collapsed).

How it was built

  • Official releases only, at a pinned version; every record of the release is kept (no sampling, no quality filtering). Rows whose text is empty after normalisation are left out and counted.
  • Normalisation (the only change to text) is kept minimal so layout survives: Unicode NFC; Windows/Mac newlines to \n; control characters other than tab and newline removed; spaces at the end of a line removed; blank lines before the first line and blank space after the last removed. Indentation, blank lines, markup, spelling, quotes and dashes are left as released. Method texts are not normalised at all.
  • Language is detected with fastText lid.176 on the first 2,000 characters; all languages are kept.
  • pre_chatgpt compares the record's date with 2022-11-30 (ChatGPT's release); text written before it is very unlikely to be AI-generated.
  • The source's other fields are kept in meta as released.

Source notes

acl_ocl

ACL OCL: The ACL Anthology Open Corpus. Rohatgi, Qin, Aw, Lim, Hu & Kan (EMNLP 2023).

One row per paper of the release's paper table (acl-publication-info.74k.v2.parquet, 73,285 papers, in table order), then one row per GROBID JSON in Full_text_JSON.tar.gz whose ACL id is not in the table (3,880 PDFs, in tar order: mostly proceedings front matter such as prefaces and committee lists, plus some papers missing from the table; the release has no metadata for them; meta.in_publication_table = false). Nothing is filtered; all languages kept. text = the paper's GROBID parse (Full_text_JSON//json//.json, S2ORC 1.0.0 format, generated 2023-01) written as plain text: the pdf_parse body_text paragraphs and then the back_matter paragraphs (acknowledgements, GROBID 'annex' = appendices) in order; whenever a paragraph's section (sec_num + section title) differs from the previous paragraph's, the heading '

' comes first as its own line; headings and paragraphs separated by a blank line. Paragraph text as released (inline citations and footnote markers kept; GROBID extraction noise not repaired). The abstract is not in text (it is in meta.abstract, the table's GROBID abstract); figure/table captions (ref_entries), the bibliography (bib_entries) and the cite/ref/equation span annotations of the JSON are left out (structured data, not running text; available in the release by ACL id). Papers whose JSON has no body or back-matter text, or that have no JSON (5,552 table papers), use the table's full_text column when it is non-empty (meta.text_from = 'table full_text'); otherwise the row has no text and is counted as dropped_empty by the builder. Result (build of 2026-10-13): 71,021 rows = 67,118 table papers with GROBID JSON text + 370 table papers with the table's full_text (243 of these are only the abstract again) + 3,533 JSON-only PDFs; 6,144 without any text (5,797 table papers, mostly without a parse, and 347 JSON-only PDFs with an empty parse). title = the table title (BibTeX, case-protecting braces kept as released, e.g. '{E}nglish'); for JSON-only rows the GROBID title. date = year and month from the table ('YYYY-MM' when the month field names a month, e.g. 'June', 'May{--}June' -> first month; else 'YYYY'); for JSON-only rows the year in the ACL id (old ids 'P19-1001' -> 2019, new ids '2020.acl-main.1' -> 2020). source_record_id = ACL id; source_url = the table url (else https://aclanthology.org/). meta = every other column of the table as released (abstract, corpus_paper_id, pdf_hash, numcitedby, url, publisher, address, year, month, booktitle, author (BibTeX author string), pages, doi, number, volume, journal, editor, isbn, ENTRYTYPE, ID, language, note) plus in_publication_table, grobid_json (tar member) and text_from; JSON-only rows keep the JSON's own metadata (title, authors, year, venue, identifiers, abstract) under grobid_metadata. Not used: the Base_JSON folder (the same 71,613 parses as single files), the two .pkl section dumps (pickles; 'v3' adds topic labels), the citation-graph parquets, the STop topic csv and the train/val sentence tars of the paper's topic-classification side task.

Licence: The release states its licence inconsistently: README text 'CC BY-NC 4.0' (linked to by-nc/4.0), README badge 'CC BY-NC 4.0' linked to by-nc-sa/4.0, YAML tag 'mit'; the strictest reading, CC BY-NC-SA 4.0, is used. The papers themselves carry the ACL Anthology's licences: CC BY-NC-SA 3.0 for material published before 2016, CC BY 4.0 from 2016 (aclanthology.org). Row licence: cc-by-nc-sa-3.0 for papers before 2016, cc-by-nc-sa-4.0 from 2016 (and when the year is unknown). Non-commercial use only, with attribution and share-alike.

aeslc

AESLC: Annotated Enron Subject Line Corpus. Zhang & Tetreault, This Email Could Save Your Life: Introducing the Task of Email Subject Line Generation (ACL 2019).

Every email of the three official splits: train 14,436, dev 1,960, test 1,906 (18,302 files in enron_subject_line//). One row per file; split = the release's folder name (train | dev | test). text = the email body as released (AESLC's preprocessed body: the authors removed quoted, forwarded and other non-author content and the headers, one sentence per line; kept exactly as in the file). source_record_id = the file name without '.subject' (<inbox|sent>). meta: subject (the original subject line), ann0, ann1, ann2 (dev and test only: subject lines written by three annotators), and from the file name mailbox, folder (inbox = received, sent = written by the mailbox owner) and msg_no. Only the AESLC release is used: no field comes from the upstream CMU Enron corpus (no headers, dates or addresses). The release has no dates or senders: date = '2003', an upper bound (the year the U.S. Federal Energy Regulatory Commission made the Enron emails public), not the date of writing; author_id = None. Nothing is left out.

Licence: CC BY-NC-SA 4.0 (AESLC README and LICENSE.md). Non-commercial use, attribution, share-alike. The underlying Enron emails were made public by the U.S. Federal Energy Regulatory Commission.

corefl

COREFL: Corpus of English as a Foreign Language. Lozano, Díaz-Negrillo & Callies, 'Designing and compiling a learner corpus of written and spoken narratives: COREFL' (in What's in a Narrative?, Peter Lang 2020); COREFL v2.0 (Oct. 2025).

Every text of both subcorpora of the release: 3,940 learner texts and 1,698 native-control texts (5,638, the total on the site's Statistics page), written texts and transcripts of the spoken versions (the same participant retold the same task orally at least 15 days later), in every language (learners write English; native controls write their L1: Spanish, English, Greek, German, ...). Split = subcorpus (learners, native). text = Text exactly as exported (verbatim: CC BY-NC-ND; spoken transcripts keep the transcription conventions). source_record_id = Filename (one Filename, DE_WR_20_14_JW in the native export, names two different texts; the second in file order gets '#2'). meta = every other column of the export as released (L1, Variety of native language, parents' languages, proficiency self-assessments, placement test score raw and %, age, sex, institution, major, years studying English, Medium, Task number, Task title, Writing/audio details, minutes taken, resources used, Original text — the original-script version of a few romanised native texts —, ...; columns differ between the two subcorpora). prompt = the task's wording from the site's Corpus design page (Tasks table): the English wording for texts in English (all learner texts and the L1-English controls), the Spanish wording for the L1-Spanish controls, and the page's bilingual description (English line, then Spanish line) for controls of other L1s, whose instruction language the release does not give; tasks 13 (Frog: retell a picture story, pictures at https://goo.gl/so3S6W) and 14 (Chaplin: retell a 4-minute video clip) have one wording, kept with its link. genre = 'learner narrative' or 'native-speaker narrative', '(written)' or '(spoken, transcribed)' from Medium. date = Year data collection; for the ranges '2014-2016' and '2006-2016' the last year (an upper bound). Nothing is left out; the audio of the spoken texts is not distributed by the site.

Licence: CC BY-NC-ND 3.0 ES (site home page and About > How to cite: 'COREFL is released under a Creative Commons license (CC BY-NC-ND 3.0 ES) (Attribution - NonCommercial - NoDerivatives)'); cite Lozano, Díaz-Negrillo & Callies (2020). Texts kept verbatim. The site's export form asks for the intended usage (we stated research use).

deceptive_opinion_spam

Deceptive Opinion Spam Corpus v1.4. Ott, Choi, Cardie & Hancock (ACL 2011); Ott, Cardie & Hancock (NAACL 2013).

Every review of the release: 1,600 .txt files (4 folders x 400: positive_polarity/deceptive_from_MTurk, positive_polarity/truthful_from_TripAdvisor, negative_polarity/deceptive_from_MTurk, negative_polarity/truthful_from_Web; 20 reviews of each of 20 Chicago hotels per folder). One row per file; text = the file content. The deceptive reviews are not AI text: they were written to order by Mechanical Turk workers (people) who had not stayed at the hotel, so all 1,600 are human-written and all are kept; the class is in meta.label (truthful | deceptive). source_record_id = the file path inside the release folder (e.g. positive_polarity/deceptive_from_MTurk/fold1/d_hilton_1.txt). meta: polarity (positive | negative), label, origin (MTurk, TripAdvisor, or Web = Expedia, Hotels.com, Orbitz, Priceline, TripAdvisor and Yelp), fold (the release's 5 cross-validation folds; not a train/test split), hotel (the file-name code) and hotel_name (from the README). The release has no dates: date = the year of the paper that released the set, an upper bound (2011 for positive, 2013 for negative reviews), not the date of writing. README.md and LICENSE are not rows. Nothing is left out.

Licence: CC BY-NC-SA 3.0 Unported (release LICENSE file and README.md: 'This work is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License'). Non-commercial use, attribution (cite the papers), share-alike.

echr

ECHR: Neural Legal Judgment Prediction in English (European Court of Human Rights cases). Chalkidis, Androutsopoulos & Aletras (ACL 2019).

Every case of the three official splits, one row per case: EN_train 7,100 cases (DATE 1960-2013), EN_dev 1,380 (1967-2013), EN_test 2,998 (2014-2018) = 11,478 (paper Table 1); the split names are the folder names without 'EN_' (train, dev, test). Each case is one JSON file .json. text = the case's TEXT list (the paragraphs of the facts section of the judgment or decision as extracted by the authors from HUDOC, with the Court's paragraph numbers such as '7. ' kept) joined with one newline, so each released paragraph is one line; 290 paragraphs of the release are empty strings and show as empty lines. Nothing else is changed. title = DOCNAME; date = DATE (the judgment/decision year, the only date released); genre = 'court judgment (statement of facts)' for BRANCH CHAMBER, GRANDCHAMBER and COMMITTEE, 'court admissibility decision (statement of facts)' for ADMISSIBILITY; source_record_id = ITEMID (HUDOC document id, unique across splits); source_url = the HUDOC page of the document. meta = every other field of the case as released: ITEMID, LANGUAGEISOCODE (ENG for all), RESPONDENT, BRANCH, DATE, DOCNAME, IMPORTANCE (HUDOC importance level, 1 = key case .. 4), CONCLUSION, JUDGES, VIOLATED_ARTICLES, VIOLATED_PARAGRAPHS, VIOLATED_BULLETPOINTS, NON_VIOLATED_ARTICLES, NON_VIOLATED_PARAGRAPHS, NON_VIOLATED_BULLETPOINTS (lists as released), plus the file name in the zip (zip_member) and n_paragraphs. Left out: the EN_train_Anon / EN_dev_Anon / EN_test_Anon folders (the same 11,478 cases again with named entities masked by the authors' spaCy run: a derived copy of the same texts) and a __MACOSX resource-fork file. The texts are English versions of the Court's documents (some written in French and translated by the Court's registry); no AI-generated text. The builder's language id marks 11 cases as uk/pl/tr/de: they are English texts dense with local names and original-language terms.

Licence: The release (archive.org item and zip) states no licence; the paper (footnote 3) says HUDOC's licensing conditions are compatible with releasing the dataset. The texts are the European Court of Human Rights' own judgments and decisions from HUDOC. The Court's reuse condition (echr.coe.int 'Copyright and Disclaimer', read 2026-10-13): "The information and texts available on the Court's website may be reproduced provided the source is acknowledged (© ECHR-CEDH) and the reproduction is made for private use or for the purposes of information and education in connection with the Court's activities. This authorisation is subject to the condition that the source is indicated and that any such reproduction is free of charge. Persons wishing to use information and texts on the Court's website for any use other than those listed above, and in particular for commercial use, must request prior written permission using the form available on the website." It also notes that the electronic version of judgments may differ from the formally adopted text. Acknowledge the source as © ECHR-CEDH (HUDOC) and cite Chalkidis et al. (ACL 2019).

ellipse

ELLIPSE: English Language Learner Insight, Proficiency and Skills Evaluation Corpus. Crossley, Tian, Baffour, Franklin, Kim, Morris, Benner, Picou & Boser (International Journal of Learner Corpus Research 9(2), 2023).

Every essay of the release, from its three data files. (1) Subset ellipse = the final corpus: split train = every row of ELLIPSE_Final_github_train.csv (3,911 essays), split test = every row of ELLIPSE_Final_github_test.csv inside the password-protected test zip (2,571; password from the README). text = full_text; source_record_id = text_id_kaggle (the release's id, unique); meta = every other column as released (gender, grade, race_ethnicity, SES, task — 'Independent' for all —, prompt, the averaged scores Overall, Cohesion, Syntax, Vocabulary, Phraseology, Grammar, Conventions, the text statistics num_words, num_words2, num_words3, num_sent, num_para, num_word_div_para, MTLD, TTR, Type, Token, and set). The release names each essay's prompt only by a short topic label (prompt column, e.g. 'Benefits of a problem'; 44 labels), not the assignment wording, so the prompt column is left empty and the label is in meta.prompt. (2) The raw rater file ellipsis_raw_rater_scores_anon_all_essay.csv (password-protected zip, password from the README; 8,890 rows: Filename, text_id_kaggle, Text, then Rater/Overall/Cohesion/Syntax/Vocabulary/Phraseology/Grammar/Conventions/Identifying_Info for rater 1 and rater 2) holds every rated essay, final or not. Each of its rows is matched to a final essay: 6,468 by text_id_kaggle; 14 whose id the release mangled into scientific notation (e.g. '6.19E+10') by identical letters-only text (lower-cased, non-letters removed) — this pairs each of the 14 final essays that had no id match with exactly one raw row; and 8 rows without an id whose letters-only text equals a final essay's (a second rating of the same essay; same Filename). These 6,490 raw rows are kept in the final essay's meta.raw_ratings (a list, all columns as released; the raw Text is included only when it differs from full_text after the corpus's text normalisation (Windows line breaks, trailing spaces) — the final texts have small cleanups, mostly a space added after sentence-final full stops — and meta.raw_ratings[i].text_same_as_final says which). (3) Subset ellipse_excluded, split full = the other 2,400 raw rows (no text_id_kaggle; essays the authors did not put in the final corpus): text = Text, meta = the other columns as released, source_record_id = Filename (with '#2' appended to a Filename's second occurrence). Nothing is left out; the rubric (ELL_Rubrics.docx) is not a text record. Anonymisation tokens (Generic_Name, STUDENT_NAME, TEACHER_NAME, ...) are kept as released. The release has no per-essay date: date = 2022-08 for every essay, an upper bound — the training essays were published, and the test essays used as the hidden test set, in Kaggle's Feedback Prize - English Language Learning (2022-08-31 to 2022-11-30), and all raw-file essays were rated before the final corpus was selected from them; essays were written in state tests (in-text years up to 2021). No author ids or URLs are released.

Licence: CC BY-NC-SA 4.0 (official README: "The data is provided under a CC BY-NC-SA 4.0 DEED Attribution-NonCommercial-ShareAlike 4.0 International license"; the repo has no LICENSE file). Non-commercial use, attribution, share-alike.

ellipse_excluded

ELLIPSE: English Language Learner Insight, Proficiency and Skills Evaluation Corpus. Crossley, Tian, Baffour, Franklin, Kim, Morris, Benner, Picou & Boser (International Journal of Learner Corpus Research 9(2), 2023).

Every essay of the release, from its three data files. (1) Subset ellipse = the final corpus: split train = every row of ELLIPSE_Final_github_train.csv (3,911 essays), split test = every row of ELLIPSE_Final_github_test.csv inside the password-protected test zip (2,571; password from the README). text = full_text; source_record_id = text_id_kaggle (the release's id, unique); meta = every other column as released (gender, grade, race_ethnicity, SES, task — 'Independent' for all —, prompt, the averaged scores Overall, Cohesion, Syntax, Vocabulary, Phraseology, Grammar, Conventions, the text statistics num_words, num_words2, num_words3, num_sent, num_para, num_word_div_para, MTLD, TTR, Type, Token, and set). The release names each essay's prompt only by a short topic label (prompt column, e.g. 'Benefits of a problem'; 44 labels), not the assignment wording, so the prompt column is left empty and the label is in meta.prompt. (2) The raw rater file ellipsis_raw_rater_scores_anon_all_essay.csv (password-protected zip, password from the README; 8,890 rows: Filename, text_id_kaggle, Text, then Rater/Overall/Cohesion/Syntax/Vocabulary/Phraseology/Grammar/Conventions/Identifying_Info for rater 1 and rater 2) holds every rated essay, final or not. Each of its rows is matched to a final essay: 6,468 by text_id_kaggle; 14 whose id the release mangled into scientific notation (e.g. '6.19E+10') by identical letters-only text (lower-cased, non-letters removed) — this pairs each of the 14 final essays that had no id match with exactly one raw row; and 8 rows without an id whose letters-only text equals a final essay's (a second rating of the same essay; same Filename). These 6,490 raw rows are kept in the final essay's meta.raw_ratings (a list, all columns as released; the raw Text is included only when it differs from full_text after the corpus's text normalisation (Windows line breaks, trailing spaces) — the final texts have small cleanups, mostly a space added after sentence-final full stops — and meta.raw_ratings[i].text_same_as_final says which). (3) Subset ellipse_excluded, split full = the other 2,400 raw rows (no text_id_kaggle; essays the authors did not put in the final corpus): text = Text, meta = the other columns as released, source_record_id = Filename (with '#2' appended to a Filename's second occurrence). Nothing is left out; the rubric (ELL_Rubrics.docx) is not a text record. Anonymisation tokens (Generic_Name, STUDENT_NAME, TEACHER_NAME, ...) are kept as released. The release has no per-essay date: date = 2022-08 for every essay, an upper bound — the training essays were published, and the test essays used as the hidden test set, in Kaggle's Feedback Prize - English Language Learning (2022-08-31 to 2022-11-30), and all raw-file essays were rated before the final corpus was selected from them; essays were written in state tests (in-text years up to 2021). No author ids or URLs are released.

Licence: CC BY-NC-SA 4.0 (official README: "The data is provided under a CC BY-NC-SA 4.0 DEED Attribution-NonCommercial-ShareAlike 4.0 International license"; the repo has no LICENSE file). Non-commercial use, attribution, share-alike.

hupd

HUPD: The Harvard USPTO Patent Dataset. Suzgun, Melas-Kyriazi, Sarkar, Kominers & Shieber (NeurIPS 2023 Datasets and Benchmarks; arXiv 2022).

Every application of the release: each JSON file in the 15 yearly archives data/.tar.gz (utility patent applications filed 2004-2018 and published by the USPTO), one row per application, read in archive order, year by year: 4,518,254 applications, exactly the creators' per-year lists (dataset-creation-utilities/lists; the paper says 4,518,263); none has an empty text. The archive year is the filing year by the creators' metadata; for 53 applications the JSON filing_date has another year (kept as released). Nothing is sampled or filtered; the release has no splits (the loader makes train/validation sets from user-chosen filing-date ranges), so split = 'full'. text = the application's abstract, claims and full description, each under a heading line ('Abstract', 'Claims', 'Description'; headings ours), separated by a blank line; an empty field is left out. The fields are as released: HUPD stores each as one line (paragraph breaks were flattened when the creators converted the USPTO XML), with the description's own section headings inline in the text ('CROSS REFERENCE TO RELATED APPLICATIONS This application ...', 'BACKGROUND ...'). HUPD's separate background and summary fields are sections HUPD extracted from the description: in a check of 9,000 applications (3,000 each of 2005, 2010 and 2017), every non-empty background and summary but 2 of 15,941 occurs in full_description once whitespace and the XML residue some of them carry ('detailed-description description="Detailed Description" end="lead"') are disregarded (the 2 differ inside an inline formula). So they are not repeated in text; they are kept unchanged in meta (with their ' HEADING ' tags). title = title; date = filing_date (YYYYMMDD in the release, here YYYY-MM-DD); genre = 'patent application'; source_record_id = application_number; source_url = the application in USPTO Patent Center (https://patentcenter.uspto.gov/applications/; formed by us, the release has no URL); author_id empty (inventors are in meta.inventor_list). meta = every other field of the JSON unchanged (application_number, publication_number, decision (ACCEPTED 2,555,166 / REJECTED 1,220,309 / PENDING 742,779, the creators' outcome label), date_produced, date_published, main_cpc_label, cpc_labels, main_ipcr_label, ipcr_labels, patent_number, filing_date, patent_issue_date, abandon_date, uspc_class, uspc_subclass, examiner_id, examiner_name_last, examiner_name_first, examiner_name_middle, inventor_list, background, summary; strings such as 'None' and 'nan' kept as released) and tar_member (the JSON's path in the archive).

Licence: Stated inconsistently by the creators: HF card metadata 'cc-by-sa-4.0'; HF card text 'HUPD is released under the CreativeCommons Attribution-NonCommercial-ShareAlike 4.0 International'; the paper (arXiv 2207.04043) 'released under the Creative Commons Attribution 4.0 International License'. Rows carry the strictest of these, CC BY-NC-SA 4.0 (non-commercial, attribution, share-alike), and so go to the non-commercial repo. The texts themselves are patent applications published by the USPTO (public record).

ice

ICE-Nigeria: the Nigerian component of the International Corpus of English. Wunder, Voormann & Gut (2010), The ICE Nigeria corpus project: Creating an open, rich and accurate corpus; ICE-Nigeria release of 2015-11-03 (SourceForge).

Every text of the release: all 510 written texts and all 392 spoken transcripts (902 rows, one per text; written first, then spoken, each by category folder and file name). text is the compilers' own plain-text version of the text from the zip, 'txt - without speaker tags/<written|spoken>//.txt'. For written texts it is byte-identical to the 'with speaker tags' copy. For spoken texts the 'with speaker tags' copy prefixes each line with the ELAN tier of its speaker ('Transcription 1 ', ...; 172 multi-speaker transcripts, line for line the same text otherwise, checked on all 392); that speaker sequence is kept in meta.speaker_lines as runs [[tier, number of lines], ...] over the lines of text (null for the 220 single-tier transcripts), so the tagged copy can be rebuilt and turns attributed, and the tiers' speaker metadata is in meta.tiers. Why the txt and not the annotated files: the txt files are the release's plain-text rendering of the same texts and are clean UTF-8, while 87 written XML files have a broken encoding (U+201D saved as e2 80 3f) and one .eaf file is not well-formed. Checked on every written file: the txt equals the XML

Licence: CC BY-NC-SA 3.0 DE (Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Germany, title page of Manual_ICE-Nigeria.pdf in the release; SourceForge lists 'Other License'). Recorded as cc-by-nc-sa-3.0 (the German port has the same terms). Non-commercial use, attribution to the ICE-Nigeria compilers (Gut et al.), share-alike.

icle_locness

ICLE500: a 500-essay open sample of the International Corpus of Learner English (ICLEv3). Granger, Dupont, Meunier, Naets & Paquot (2020), ICLEv3; ICLE500 data release, UCLouvain Dataverse (2024), doi:10.14428/DVN/RIOSSC.

All 500 essays of ICLE500 (the release's Texts/ folder: 500 .txt files, one essay each; 20 per L1 group x 25 groups), one row per essay, in file-name order. Each file starts with a line 'Title: '; <code>title</code> = that line without 'Title:' (the essay title as given in the ICLE file, usually the prompt in the student's or the compilers' wording) and <code>text</code> = the rest of the file (the learner's essay, errors and ICLE editorial tokens '[Quotation]', '[Reference]', '[Illegible word/phrase]' and '[?..?]' (uncertain reading) kept as released). <code>prompt</code> = the <code>prompt</code> column of ICLE_rubric_rated_500.csv (the prompt the essay was written in response to). <code>source_record_id</code> = the ICLE text id (<code>filename</code> column, e.g. SPM04048; the released file name, numbered and sometimes suffixed '_relabelled', is meta.dataverse_file). <code>meta</code> holds the essay's row of ICLE_rubric_rated_500.csv as released: <code>filename</code>, <code>L1</code>, <code>prompt</code>, <code>length_in_words</code>, <code>fair_mean_avge_MFRM</code>, <code>measure_MFRM</code>, <code>rank_rubric_MFRM</code>, <code>CEFR_MFRM</code> (Rasch fair-mean average, measure, rank and CEFR band from the rubric-based assessment; definitions in the xlsx sheet 'column_names'), plus <code>dataverse_file</code>. The release has no per-essay date: <code>date</code> is 2020 for every essay, an upper bound (ICLEv3, from which the sample is drawn, was published in 2020), not the year of writing. No author ids or per-essay URLs. Not texts and left out: the two metadata spreadsheets (used for meta), the metadata documentation PDF, the technical report PDF and readme.txt. Nothing else is left out. The full ICLEv3 (>9,000 essays, paid licence) and LOCNESS (licence form) are not openly released and are not included.</p> <p><em>Licence:</em> CC BY-NC (Dataverse terms of use of the dataset: 'CC-BY-NC'; no version given, CC BY-NC 4.0 assumed). Non-commercial use with attribution to the ICLE team (Granger et al.) and the ICLE500 data release.</p> <h3 class="relative group flex items-baseline"> <a id="idealens" class="block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full" href="#idealens" > <span class="header-link"><!--[--><svg class="text-gray-500 hover:text-black dark:hover:text-gray-200 w-4" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!--]--></span> </a> <span> <code>idealens</code> </span> </h3> <p>TwiceTold (IdeaLens): human stories written from AI outlines. Rajendhran, Choi, Russell, Namuduri, Bölöni-Turgut, Karpinska, Wieting & Iyyer (arXiv 2026).</p> <p>The release has two splits: human_rewrite (50 stories that 17 computer-science graduate students and faculty, ids W01-W17, wrote in person, each from the outline of an AI-generated story) and ai_original (the 50 GPT-5.6 Sol stories those outlines were extracted from). Included: all 50 human_rewrite stories (split human_rewrite). Left out: the 50 ai_original stories (AI-generated prose). The ideas and plot of every included story are the AI's; the prose is human (card: 'The prose of the 50 rewrites is human'). <code>source_record_id</code> = <code>id</code> (twicetold_NN_human); <code>text</code> = <code>text</code>; <code>prompt</code> is empty: the brief each writer worked from is <code>writer_outline</code> (an ordered list of '[Role] content' items extracted by a model from the AI story), which was written by an AI model, so it is kept only in meta, as <code>outline_ai_written</code>, with <code>outline_note</code> saying so; <code>genre</code> = 'short story (<genre>)'; <code>author_id</code> = <code>writer</code>. <code>date</code> = 2026: the outlines come from GPT-5.6 Sol stories (system card July 2026) and the data were first committed on 2026-09-29, so the stories were written Jul-Sep 2026. <code>meta</code> holds every released field except the text: id, story (pair number shared with the left-out AI story), kind, prose_author, idea_author, writer, genre, generation_prompt (the prompt that produced the AI story), outline_ai_written (= the release's writer_outline, renamed), words, and the detector scores reported in the paper (idealens_p_ai, proselens_p_ai, editlens_score, pangram4_p_ai).</p> <p><em>Licence:</em> CC BY-NC-SA 4.0 (dataset card and LICENSE file: 'The dataset, including the human-written stories, is released under CC BY-NC-SA 4.0'). The card notes that the original AI stories and outlines were produced with OpenAI models.</p> <h3 class="relative group flex items-baseline"> <a id="keystroke_learner_writing" class="block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full" href="#keystroke_learner_writing" > <span class="header-link"><!--[--><svg class="text-gray-500 hover:text-black dark:hover:text-gray-200 w-4" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!--]--></span> </a> <span> <code>keystroke_learner_writing</code> </span> </h3> <p>KUPA-KEYS: Logging Keystrokes in Writing by English Learners. Velentzas, Caines, Borgo, Pacquetet, Hamilton, Arnold, Nicholls, Buttery, Gaillat, Ballier & Yannakoudakis (LREC-COLING 2024).</p> <p>Every row of the three data files; the release has no splits (split <code>full</code>). (1) <code>keystroke_learner_writing</code> (documents): the 1,006 rows of KUPA-KEYS-META.csv in file order, one per participant. <code>text</code> = <code>answer</code> (the final essay as submitted in the logged textbox); <code>prompt</code> = <code>prompt</code> (the W&I 'Just for Fun' prompt the participant was randomly given, 10 prompts); <code>source_record_id</code> = <code>author_id</code> = <code>id</code> (the anonymised Qualtrics submission id, one essay per participant). <code>meta</code> = all other META columns as released (strings; empty cells as ''): navigator_useragent, navigator_language, age, handedness, comphours, layoutcomf, layoutnow, comptype, countryres, nativelang, englishyears, englishcountrymonths, englishexposure, otherlanguages, cefrself, cefrlevel, cefrwhen, educ, time, task1_time, task2_time, task1_events, task2_events, task1_words, task2_words, cohort, prompt_id, mark_a0, mark_h1-h3 (W&I automarker and three human marks 0-12), cefr_a0, cefr_h1-h3. <code>date</code> = 2023: no collection date is released; the three Prolific cohorts ran in 2023 (paper: submissions screened with 'the tool released by OpenAI on January 31st, 2023'; the browser versions in navigator_useragent are mostly Chrome 109-117 / Firefox 108-114, released January-September 2023; HF release March 2024). Native English speakers (nativelang 'English', cohort 1) are included like everyone else. (2) <code>keystroke_learner_writing_keystrokes</code> (logs): 2,012 rows, one per participant and task: the participant's rows of KUPA-KEYS-TASK-2.csv (essay writing; <code>source_record_id</code> '<id>/task2') and of KUPA-KEYS-TASK-1.csv (text-copy task; '<id>/task1'), format <code>csv</code>: the file's header line followed by that participant's records, byte for byte as in the file (UTF-8; LF line ends; quoted multi-line 'capture' and 'input' text kept), in file order. <code>parent_id</code> = the participant's essay row id 'keystroke_learner_writing/<id>'; meta = task, file, participant id, number of events. The copy task's text is a given text retyped, not the participant's own writing, so it is kept only as a log, not as a document (META has no copy-task text). Event columns: id, time, type (down, up, click, input, capture), key, key_code, alt_key, ctrl_key, meta_key, shift_key, is_repeat, range_start, range_end, text (DatasetDescription.md). Nothing is sampled or filtered; README.md and DatasetDescription.md are documentation, not records.</p> <p><em>Licence:</em> CC BY-NC-SA 4.0 (HF card 'license: cc-by-nc-sa-4.0'). Non-commercial use, attribution, share-alike. The W&I prompts are 'used with permission' of ELiT (README).</p> <h3 class="relative group flex items-baseline"> <a id="keystroke_learner_writing_keystrokes" class="block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full" href="#keystroke_learner_writing_keystrokes" > <span class="header-link"><!--[--><svg class="text-gray-500 hover:text-black dark:hover:text-gray-200 w-4" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!--]--></span> </a> <span> <code>keystroke_learner_writing_keystrokes</code> </span> </h3> <p>KUPA-KEYS: Logging Keystrokes in Writing by English Learners. Velentzas, Caines, Borgo, Pacquetet, Hamilton, Arnold, Nicholls, Buttery, Gaillat, Ballier & Yannakoudakis (LREC-COLING 2024).</p> <p>Every row of the three data files; the release has no splits (split <code>full</code>). (1) <code>keystroke_learner_writing</code> (documents): the 1,006 rows of KUPA-KEYS-META.csv in file order, one per participant. <code>text</code> = <code>answer</code> (the final essay as submitted in the logged textbox); <code>prompt</code> = <code>prompt</code> (the W&I 'Just for Fun' prompt the participant was randomly given, 10 prompts); <code>source_record_id</code> = <code>author_id</code> = <code>id</code> (the anonymised Qualtrics submission id, one essay per participant). <code>meta</code> = all other META columns as released (strings; empty cells as ''): navigator_useragent, navigator_language, age, handedness, comphours, layoutcomf, layoutnow, comptype, countryres, nativelang, englishyears, englishcountrymonths, englishexposure, otherlanguages, cefrself, cefrlevel, cefrwhen, educ, time, task1_time, task2_time, task1_events, task2_events, task1_words, task2_words, cohort, prompt_id, mark_a0, mark_h1-h3 (W&I automarker and three human marks 0-12), cefr_a0, cefr_h1-h3. <code>date</code> = 2023: no collection date is released; the three Prolific cohorts ran in 2023 (paper: submissions screened with 'the tool released by OpenAI on January 31st, 2023'; the browser versions in navigator_useragent are mostly Chrome 109-117 / Firefox 108-114, released January-September 2023; HF release March 2024). Native English speakers (nativelang 'English', cohort 1) are included like everyone else. (2) <code>keystroke_learner_writing_keystrokes</code> (logs): 2,012 rows, one per participant and task: the participant's rows of KUPA-KEYS-TASK-2.csv (essay writing; <code>source_record_id</code> '<id>/task2') and of KUPA-KEYS-TASK-1.csv (text-copy task; '<id>/task1'), format <code>csv</code>: the file's header line followed by that participant's records, byte for byte as in the file (UTF-8; LF line ends; quoted multi-line 'capture' and 'input' text kept), in file order. <code>parent_id</code> = the participant's essay row id 'keystroke_learner_writing/<id>'; meta = task, file, participant id, number of events. The copy task's text is a given text retyped, not the participant's own writing, so it is kept only as a log, not as a document (META has no copy-task text). Event columns: id, time, type (down, up, click, input, capture), key, key_code, alt_key, ctrl_key, meta_key, shift_key, is_repeat, range_start, range_end, text (DatasetDescription.md). Nothing is sampled or filtered; README.md and DatasetDescription.md are documentation, not records.</p> <p><em>Licence:</em> CC BY-NC-SA 4.0 (HF card 'license: cc-by-nc-sa-4.0'). Non-commercial use, attribution, share-alike. The W&I prompts are 'used with permission' of ELiT (README).</p> <h3 class="relative group flex items-baseline"> <a id="l1_signals" class="block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full" href="#l1_signals" > <span class="header-link"><!--[--><svg class="text-gray-500 hover:text-black dark:hover:text-gray-200 w-4" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!--]--></span> </a> <span> <code>l1_signals</code> </span> </h3> <p>Can We Still Trace L1 Signals? (academic abstracts labelled with the authors' native language). Utami & Sasano, 'Can We Still Trace L1 Signals? Investigating the Resilience of Native Language Signals in the LLM Era' (EMNLP 2026).</p> <p>Every row of every file of the release (2,800 abstracts), one split per file: acl_prenn (400, years 1979-2015), acl_prellm (400, 2016-2022), acl_postllm (400, 2023-2025) and arxiv_set (1,600; the paper's training set). <code>text</code> = <code>abstract</code>, <code>title</code> = <code>title</code>, both as released (verbatim, since some rows are CC BY-NC-ND); <code>date</code> = <code>year</code> (ACL files) or <code>publication_year</code> (arXiv file; its <code>year</code> is null); <code>source_record_id</code> = '<file>:<0-based line>' (the release has no id: DOIs repeat or are missing); <code>source_url</code> = <code>link</code> if given, else <a href="https://doi.org/" rel="nofollow">https://doi.org/</a><doi>; <code>meta</code> = doi, year, publication_year, link, authors (names, OpenAlex institutions and, in the arXiv file, the predicted country, as released) and native_lang (the L1 group, inferred by the release authors from names and affiliations: a noisy label). The release repeats some abstracts (3 identical pairs in acl_prenn; in acl_postllm some papers appear up to 5 times, once per author-L1 label); every row is kept. Licence per row, by where the abstract was published: ACL Anthology material (DOI 10.3115/ or 10.18653/, an aclanthology.org link, or a TACL article 10.1162/tacl) follows the Anthology's terms: published in or after 2016 cc-by-4.0, before 2016 cc-by-nc-sa-3.0 (for TACL, whose own licence is CC BY 4.0, the stricter Anthology term decides before 2016). Computational Linguistics articles (other 10.1162/ DOIs) are published by MIT Press under CC BY-NC-ND 4.0, stricter than the Anthology's statement: cc-by-nc-nd-4.0. The arXiv training file and acl_postllm rows linked to arxiv.org are arXiv abstracts: arXiv metadata is CC0 (cc0-1.0). Abstracts of papers outside the ACL policy that state no open licence — IJCNLP 2005 (Springer, DOI 10.1007/), RANLP (DOI 10.26615/), and acl_postllm rows linked to CVF open access (CVPR/ICCV), the ISCA archive (Interspeech 2024), NeurIPS proceedings or personal pages — carry <code>quotation</code> (a short verbatim quote of a paper, non-commercial research use with credit). Nothing is left out.</p> <p><em>Licence:</em> The repository states no licence (no LICENSE file; the README gives none). The abstracts belong to their papers: ACL Anthology materials are CC BY 4.0 from 2016 and CC BY-NC-SA 3.0 before (aclanthology.org/faq/copyright: 'other materials are copyrighted by their respective copyright holders'); Computational Linguistics is CC BY-NC-ND 4.0, TACL CC BY 4.0 (MIT Press / DOAJ); arXiv metadata including abstracts is CC0; abstracts of papers with no open licence are shared as quotations for non-commercial research with credit to the paper. Author names and affiliations in meta come from OpenAlex (CC0). The L1 labels are the release authors' inferences.</p> <h3 class="relative group flex items-baseline"> <a id="methods" class="block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full" href="#methods" > <span class="header-link"><!--[--><svg class="text-gray-500 hover:text-black dark:hover:text-gray-200 w-4" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!--]--></span> </a> <span> <code>methods</code> </span> </h3> <p>anagram-bench generation methods. anagram-bench (2026); each row cites the paper or repository it quotes.</p> <p>The 22 task recipes of the anagram-bench build of 2026-10-09 list how each task's AI text is made: prompt templates, instruction pools and attacks, each traced to a paper, its code or a tool (with URL, revision and location). Every template is a row (kind <code>template</code>, or <code>classifier</code> for judge/filter prompts such as pv_rw0305_compliance_judge or mine_classifier_v1); every item of an instruction pool is a row (kind <code>instruction</code>, <code>name</code> = pool id, <code>meta.item_index</code> = position in the pool); every attack is a row (kind <code>attack</code>, <code>text</code> null, <code>params</code> = its settings as JSON, <code>meta</code> = code_url, revision, entry_point). A template used by two tasks appears once per task (<code>task</code>). <code>text</code> and <code>system</code> are stored exactly as in the recipe (no normalisation); a template stored as a list of turns (mirror_pangram1_title_then_essay) is joined with a blank line and the list kept in <code>meta.text_turns</code>. <code>verbatim</code> is the recipe's flag (instruction items: true unless the pool is our own or a paraphrase of API documentation). <code>paper</code>/<code>paper_url</code> come from the build's work list (works_813.json); <code>location</code> is the recipe's provenance location (attacks: the entry point). Template settings (params, generation_params) go to <code>params</code>; all other recipe fields go to <code>meta</code> unchanged; pool-level fields go to every item's <code>meta</code>, per-item fields (item_meta, item_ids, ...) only to their item. LEFT OUT: instruction pools that are user data rather than paper methods, which live in their own datasets: rw0181_wildchat_second_turn_followups (follow-up requests mined from WildChat-1M (allenai/WildChat-1M); pointer only in the recipe (no items)); rw0181_wildchat_1m_edit_request_candidates (9,673 edit requests mined from WildChat-1M user turns (allenai/WildChat-1M)); new0025_lmsys_chat_1m_edit_request_candidates (5,968 edit requests mined from LMSYS-Chat-1M user turns (lmsys/lmsys-chat-1m)); new0259_kiwi_revision_instructions (KIWI revision instructions written by users (fangyuan/kiwi); pointer only in the recipe (no items)); new0008_recipe4u_request_for_revision (RECIPE4U student requests (consent-form access, zeunie.github.io/RECIPE4U); pointer only (no items)); new0024_openrewriteeval_instructions (1,033 human-written rewrite instructions of OpenRewriteEval (google-research/rewritelm)); new0163_no_robots_rewrite_test_instructions (35 annotator-written Rewrite requests of No Robots (HuggingFaceH4/no_robots; in the corpus as subset no_robots)); new0035_nirvana_editing_queries (32 student queries to ChatGPT from NIRVANA (OSF 3a8uh)); lad_pool_longbench_write_en_ge1500 (24 LongBench-Write user prompts (THUDM/LongWriter, longbench_write.jsonl)). LICENCES (per row; <code>meta.licence_basis</code> says why): text quoted from a paper takes the paper's arXiv licence (OAI-PMH record) or CC BY 4.0 if the paper is in the ACL Anthology (2016 on), whichever is more permissive; arXiv's non-exclusive licence alone gives <code>quotation</code> (non-commercial repo). Text quoted from code or data takes the repository's licence (GitHub / Hugging Face; LICENSE files read by hand where GitHub reports none or NOASSERTION), <code>quotation</code> if there is none, unless the paper prints the same wording. Our own wrappers are CC BY 4.0; templates that hold only placeholders, attack settings, and lists of names or numbers are <code>facts</code>. Decided by the corpus owner: EditLens CC BY-NC-SA 4.0, Pangram 4 Technical Report CC BY-NC-ND 4.0, OpenTuringBench CC BY-NC 4.0. EMBEDDED PASSAGES: the long templates were read; where a template embeds an example passage it is named in <code>meta.embedded_source</code> and the row takes the stricter licence. Two RW0302 chain-of-thought templates embed a CNN/DailyMail article and an Amazon review, which cannot be redistributed: those passages are replaced by {EXAMPLE_REMOVED} (listed in <code>meta.removed_passages</code>, sha256 of the original text in <code>meta.original_text_sha256</code>). PAID SERVICES (StealthGPT, Undetectable AI, Humbot, BypassGPT, QuillBot, DAMAGE's web tools) are given by name and settings only, <code>meta.note</code> = 'paid service; not run': the vendors' descriptions of their modes are cut from the mode lists (rw0389_quillbot_paraphraser_modes, rw0278_vendor_modes: <code>text</code> = the mode name, <code>meta.vendor_description_removed</code>) and from the Humbot/BypassGPT attack settings (<code>meta.params_trimmed</code>), and rw0770_leaked_humanizer_prompt_fragments (3 fragments of commercial humanizers' system prompts revealed by jailbreak, quoted in DAMAGE (RW0770) §3.2 and App. A) is left out.</p> <p><em>Licence:</em> Per row, in <code>license</code>: the licence of the quoted paper, repository or dataset, <code>quotation</code> for text whose source carries no open licence (non-commercial research use with attribution), <code>facts</code> for settings, and CC BY 4.0 for anagram-bench's own wording. Cite the paper in <code>paper</code> / <code>paper_url</code> when using a row.</p> <h3 class="relative group flex items-baseline"> <a id="moviesum" class="block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full" href="#moviesum" > <span class="header-link"><!--[--><svg class="text-gray-500 hover:text-black dark:hover:text-gray-200 w-4" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!--]--></span> </a> <span> <code>moviesum</code> </span> </h3> <p>MovieSum: An Abstractive Summarization Dataset for Movie Screenplays. Saxena & Keller (Findings of ACL 2024).</p> <p>Every movie of every split of the official release: train 1,800 (train.jsonl), validation 200 (val.jsonl), test 200 (test.jsonl); one row per movie, in file order. <code>text</code> = <code>script</code>, the screenplay exactly as released: the authors' XML rendering (<script>, <scene>, <stage_direction>, <scene_description>, <character>, <dialogue>) of scripts collected from scriptslug.com, imsdb.com and dailyscript.com, tokenised as in the release (PTB tokens such as -LRB- and `` '' are kept). <code>title</code> = <code>movie_name</code> without its '_<year>' suffix; <code>date</code> = that year (the film's release year; the screenplay was written before or around it); <code>source_record_id</code> = <code>movie_name</code> (unique in the release; <code>imdb_id</code> is not: 12 films occur twice, as two different script files under two names, e.g. 'Mulholland Dr._2001' and 'Mulholland Drive_2001', 3 of them across train and test; both kept as released); <code>source_url</code> = the IMDb page of <code>imdb_id</code>. <code>meta</code> = movie_name, imdb_id and <code>summary</code>, the movie's Wikipedia plot summary as released (the summarisation target). Nothing is left out.</p> <p><em>Licence:</em> CC BY-NC 4.0 (dataset card: 'Creative Commons Attribution Non Commercial 4.0'; YAML license cc-by-nc-4.0). Non-commercial use with attribution. The screenplays remain the copyright of their writers and studios (the release distributes them for research), and the plot summaries in meta are Wikipedia text (CC BY-SA), redistributed by the release under its licence.</p> <h3 class="relative group flex items-baseline"> <a id="no_robots" class="block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full" href="#no_robots" > <span class="header-link"><!--[--><svg class="text-gray-500 hover:text-black dark:hover:text-gray-200 w-4" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!--]--></span> </a> <span> <code>no_robots</code> </span> </h3> <p>No Robots. Rajani, Tunstall, Beeching, Lambert, Rush & Wolf (Hugging Face H4, 2023).</p> <p>Every row of both official splits (train 9,500, test 500; the <code>*_sft</code> copies of the same rows are not repeated). One prompt id occurs twice in train with different content; the second copy's id ends in <code>#2</code>. Prompts and answers were both written by human annotators, so <code>messages</code> holds the whole exchange and <code>human_response</code> the first answer. <code>prompt</code> is the first user turn; where the release's own <code>prompt</code> field differs (it holds the system prompt for Chat rows), it is kept in <code>meta.release_prompt</code>. <code>category</code> is the release's task category (Generation, Open QA, Brainstorm, Chat, Rewrite, Summarize, Coding, Classify, Closed QA, Extract).</p> <p><em>Licence:</em> CC BY-NC 4.0 (dataset card).</p> <h3 class="relative group flex items-baseline"> <a id="pelic" class="block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full" href="#pelic" > <span class="header-link"><!--[--><svg class="text-gray-500 hover:text-black dark:hover:text-gray-200 w-4" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!--]--></span> </a> <span> <code>pelic</code> </span> </h3> <p>PELIC: University of Pittsburgh English Language Institute Corpus. Juffs, Han & Naismith (2020), The University of Pittsburgh English Language Corpus (PELIC), doi:10.5281/zenodo.3991977.</p> <p>Every row of corpus_files/answer.csv (46,204 texts; the README table still quotes v1.1's 46,230, before the 2021-03 removal of 3 texts with more than 3 versions): one row per text, in file order, including every revised version. <code>source_record_id</code> = <code>answer_id</code>; <code>text</code> = <code>text</code> exactly as released (the licence is NoDerivatives, so the text is not normalised; the corpus authors had already ASCII-fied punctuation and replaced names, URLs and e-mails with ANON_* tokens). Versions: the release codes drafts as <code>version</code> 1-3 (later versions are revisions after teacher feedback) but has no document id, so <code>parent_id</code> = '<anon_id>/<question_id>/<course_id>' (student x task x course; the README's unit of a revised assignment), the same for all versions of one assignment, and meta.version, meta.n_versions_in_group and meta.prev_answer_id (the answer_id of the latest lower version in the group by created_date, or null) link the drafts; the README warns that version coding is imperfect (some 'version 1' texts are re-submissions), so these links are as released, not checked. <code>prompt</code> = the task stem from question.csv (question_id join; 104 texts whose question_id (9 ids) has no row in question.csv get no prompt and meta.question = null). <code>author_id</code> = <code>anon_id</code>; <code>date</code> = <code>created_date</code> (ISO date-time as released); <code>genre</code> = 'learner <question type> (<class> class)' from the README glossary. <code>meta</code> holds every other answer.csv field as released: <code>answer_id</code>, <code>question_id</code>, <code>anon_id</code>, <code>course_id</code>, <code>version</code>, <code>created_date</code>, <code>text_len</code> (the release's RE token count), <code>tokens</code> (the authors' NLTK tokenisation) and <code>tok_lem_POS</code> (their automatic token/lemma/POS layer); then the question row (<code>question_type_id</code>, <code>allow_text</code>; the stem is in <code>prompt</code>), the course row (<code>class_id</code>, <code>level_id</code>, <code>semester</code>, <code>section</code>) plus the README's labels for class, question type, level and CEFR, the student's row of student_information.csv (all 21 columns as released, including course_history) and the student's placement test rows of test_scores.csv (<code>test_scores</code>: a list, usually one per entry semester). All values are strings as released. Left out: PELIC_compiled.csv (the release's join of these same files); corpus_stats, tutorials and README images (no texts). The spoken data announced in the README is not released. No text is dropped except empty ones, which the builder counts.</p> <p><em>Licence:</em> CC BY-NC-ND 4.0 (README section 11: 'licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License'; the Zenodo record says cc-by-nd-4.0, the stricter reading with NC is used). Non-commercial use, attribution to Juffs, Han & Naismith, no derivatives: share the texts only unmodified.</p> <h3 class="relative group flex items-baseline"> <a id="persuade_2" class="block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full" href="#persuade_2" > <span class="header-link"><!--[--><svg class="text-gray-500 hover:text-black dark:hover:text-gray-200 w-4" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!--]--></span> </a> <span> <code>persuade_2</code> </span> </h3> <p>PERSUADE 2.0. Crossley, Baffour, Tian, Franklin, Benner & Boser (Assessing Writing 2024).</p> <p>Every essay of both official files: train 15,594 essays (persuade_corpus_2.0_train.csv) and test 10,402 (persuade_corpus_2.0_test.csv in the password-protected test zip; password from the README). The CSVs have one row per discourse element (173,266 and 112,117 rows); a corpus row is one essay (<code>essay_id_comp</code>, the release's essay id, is <code>source_record_id</code>), in the order the essays first appear in the file. <code>text</code> = <code>full_text</code>; <code>prompt</code> = <code>assignment</code> (the assignment text the student answered; given for every essay). <code>meta</code> holds the essay-level fields as released (<code>essay_id</code> — a number mangled to scientific notation in the release, e.g. 5.40889E+12, so not unique —, <code>competition_set</code>, <code>holistic_essay_score</code>, <code>essay_word_count</code>, <code>provider</code>, <code>task</code>, <code>prompt_name</code>, <code>source_text</code> (the sources of source-based prompts), <code>gender</code>, <code>grade_level</code>, <code>ell_status</code>, <code>race_ethnicity</code>, <code>economically_disadvantaged</code>, <code>student_disability_status</code>; empty strings kept as released) and <code>discourse</code>: the essay's discourse-element rows in file order, each with <code>discourse_id</code>, <code>discourse_start</code>, <code>discourse_end</code>, <code>discourse_text</code>, <code>discourse_type</code>, <code>discourse_type_num</code>, <code>discourse_effectiveness</code>, <code>hierarchical_id</code>, <code>hierarchical_text</code>, <code>hierarchical_label</code> (strings as released). The release has no per-essay date: <code>date</code> is 2021 for every essay, an upper bound, not the year of writing — the training essays were published for the Kaggle Feedback Prize 2021 (Dec 2021) and the PERSUADE 1.0 paper (Assessing Writing 54, received 2022-02-08) already describes all 25,996 essays. No author ids or URLs are released. Nothing is left out.</p> <p><em>Licence:</em> CC BY-NC-SA 4.0 (official README: "The data is provided under a CC BY-NC-SA 4.0 DEED Attribution-NonCommercial-ShareAlike 4.0 International license"). Non-commercial use, attribution, share-alike. Disability status means the student has an IEP or 504 plan (README).</p> <h3 class="relative group flex items-baseline"> <a id="sharechat" class="block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full" href="#sharechat" > <span class="header-link"><!--[--><svg class="text-gray-500 hover:text-black dark:hover:text-gray-200 w-4" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!--]--></span> </a> <span> <code>sharechat</code> </span> </h3> <p>ShareChat: A Dataset of Chatbot Conversations in the Wild. Yan, Nguyen, Su, Lieffers & Le (arXiv 2026).</p> <p>Every conversation of the five released platform files (HF configs chatgpt, claude, gemini, grok, perplexity; each has a single split <code>train</code>, kept as the split). The files hold one row per message; a conversation is one shared link (<code>url</code>), and its messages are ordered by <code>message_index</code>. Each conversation is one row with the whole conversation in <code>messages</code> (role <code>user</code> stays user, <code>llm</code> becomes assistant; the chatbots' replies are kept, the licence covers them). Texts are the released <code>plain_text</code>, already PII-redacted by the authors (placeholders such as <REDACTED>, <URL>, <PERSON>). A few conversations repeat messages (e.g. some Perplexity threads list each message twice or appear twice in the file); they are kept as released. Empty messages stay as empty turns; 97 ChatGPT conversations start with the chatbot's message, and 16 ChatGPT conversations have no non-empty user message, so the builder drops them (129,584 conversations released: ChatGPT 94,086, Perplexity 14,612, Grok 13,149, Gemini 6,826, Claude 911). <code>source_record_id</code> and <code>source_url</code> = the conversation's url; <code>model</code> = the platform. <code>date</code> = the time of the first user message where the release gives one: ChatGPT <code>message_create_time</code> (unix time, else the conversation's <code>create_time</code>), Grok <code>message_create_time</code>, Gemini <code>created_at</code> (else <code>published_at</code>), Perplexity <code>last_updated</code> (date the thread was last updated: an upper bound); Claude has no dates. <code>meta.platform</code>, <code>meta.url</code> and <code>meta.messages</code> (per message, every other released column unchanged: message_index, role, turns_count, detected_language_final, topic and the platform's own columns such as model, create_time, thinking, code, analysis, version, links, source, source_bar, views, shares, other_info). Not included: filtered_out_conversations_non_target_languages.json, which lists only the urls, turn counts and languages of conversations the authors removed (no text).</p> <p><em>Licence:</em> CC BY-NC 4.0 (ShareChat Dataset License Agreement on the dataset card, accepted through the card's gate; DATA_LICENSE in github.com/raye22/ShareChat), with additional terms: do not attempt to identify the individuals whose conversations appear in the dataset, or to infer or recover personal data the de-identification removed; when using the outputs of a specific chatbot, also follow that platform's terms of use (ChatGPT, Claude, Gemini, Grok, Perplexity); do not use the data to build production systems whose primary purpose is to surveil, profile or de-anonymize individuals. The licence allows copying and redistribution; the card (update of 21 Apr 2026) allows sharing derivative subsets with attribution to ShareChat. Non-commercial use only.</p> <h2 class="relative group flex items-baseline"> <a id="licences-and-attribution" class="block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full" href="#licences-and-attribution" > <span class="header-link"><!--[--><svg class="text-gray-500 hover:text-black dark:hover:text-gray-200 w-4" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!--]--></span> </a> <span> Licences and attribution </span> </h2> <p>Each row keeps the licence of its source (column <code>license</code>; the table above and <code>LICENSE.md</code> list them). If you redistribute rows, credit the source as its licence requires (title, authors, link, licence) and keep share-alike rows under the same licence. Non-commercial rows may not be used for commercial purposes; rows marked BY-NC-ND may only be shared unaltered.</p> <p>The columns this corpus adds (ids, language, counts, hashes, <code>pre_chatgpt</code>, genre labels) are released under CC0. The build code is Apache-2.0.</p> <h2 class="relative group flex items-baseline"> <a id="personal-data-and-removal-requests" class="block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full" href="#personal-data-and-removal-requests" > <span class="header-link"><!--[--><svg class="text-gray-500 hover:text-black dark:hover:text-gray-200 w-4" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!--]--></span> </a> <span> Personal data and removal requests </span> </h2> <p>The texts come from public releases and may name real people. Do not try to identify people who are not public figures. If you want a text removed, open a discussion on this repository with the row id; we remove it in the next version.</p> <h2 class="relative group flex items-baseline"> <a id="citation" class="block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full" href="#citation" > <span class="header-link"><!--[--><svg class="text-gray-500 hover:text-black dark:hover:text-gray-200 w-4" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" aria-hidden="true" role="img" width="1em" height="1em" preserveAspectRatio="xMidYMid meet" viewBox="0 0 256 256"><path d="M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z" fill="currentColor"></path></svg><!--]--></span> </a> <span> Citation </span> </h2> <p>Please cite the original source of every subset you use (see the source notes). A citation for this corpus will follow with the anagram-bench paper.</p>