README.md

---
pretty_name: Anagram Corpus
license: other
license_name: anagram-corpus-mixed
license_link: LICENSE.md
language:
- en
- multilingual
tags:
- ai-text-detection
- human-written-text
- prompts
configs:
- config_name: ai_text_not_mine
  data_files:
  - split: test
    path: data/ai_text_not_mine/test-*.parquet
- config_name: ai_text_not_mine_keystrokes
  data_files:
  - split: test
    path: data/ai_text_not_mine_keystrokes/test-*.parquet
- config_name: american_stories
  data_files:
  - split: full
    path: data/american_stories/full-*.parquet
- config_name: arxiv
  data_files:
  - split: full
    path: data/arxiv/full-*.parquet
- config_name: asap_2
  data_files:
  - split: train
    path: data/asap_2/train-*.parquet
  - split: test
    path: data/asap_2/test-*.parquet
- config_name: caselaw_access_project
  data_files:
  - split: full
    path: data/caselaw_access_project/full-*.parquet
- config_name: casimir
  data_files:
  - split: train
    path: data/casimir/train-*.parquet
  - split: test
    path: data/casimir/test-*.parquet
  - split: validation
    path: data/casimir/validation-*.parquet
- config_name: casimir_reviews
  data_files:
  - split: full
    path: data/casimir_reviews/full-*.parquet
- config_name: chatbot_arena_140k
  data_files:
  - split: train
    path: data/chatbot_arena_140k/train-*.parquet
- config_name: chatbot_arena_55k
  data_files:
  - split: train
    path: data/chatbot_arena_55k/train-*.parquet
- config_name: chatgpt_detector_bias
  data_files:
  - split: full
    path: data/chatgpt_detector_bias/full-*.parquet
- config_name: civil_comments
  data_files:
  - split: train
    path: data/civil_comments/train-*.parquet
  - split: test
    path: data/civil_comments/test-*.parquet
  - split: validation
    path: data/civil_comments/validation-*.parquet
- config_name: cmu_movie_summaries
  data_files:
  - split: full
    path: data/cmu_movie_summaries/full-*.parquet
- config_name: co_storm_wildseek
  data_files:
  - split: full
    path: data/co_storm_wildseek/full-*.parquet
- config_name: common_pile_foodista
  data_files:
  - split: train
    path: data/common_pile_foodista/train-*.parquet
- config_name: common_pile_github_archive
  data_files:
  - split: train
    path: data/common_pile_github_archive/train-*.parquet
- config_name: common_pile_regulations
  data_files:
  - split: train
    path: data/common_pile_regulations/train-*.parquet
- config_name: common_pile_ubuntu_irc
  data_files:
  - split: train
    path: data/common_pile_ubuntu_irc/train-*.parquet
- config_name: common_pile_usgpo
  data_files:
  - split: train
    path: data/common_pile_usgpo/train-*.parquet
- config_name: congressional_hearings
  data_files:
  - split: full
    path: data/congressional_hearings/full-*.parquet
- config_name: congressional_record
  data_files:
  - split: full
    path: data/congressional_record/full-*.parquet
- config_name: courtlistener
  data_files:
  - split: full
    path: data/courtlistener/full-*.parquet
- config_name: cuad
  data_files:
  - split: train
    path: data/cuad/train-*.parquet
  - split: test
    path: data/cuad/test-*.parquet
- config_name: djinni
  data_files:
  - split: train
    path: data/djinni/train-*.parquet
- config_name: ebg_stylometry
  data_files:
  - split: full
    path: data/ebg_stylometry/full-*.parquet
- config_name: ebg_stylometry_bg
  data_files:
  - split: full
    path: data/ebg_stylometry_bg/full-*.parquet
- config_name: edgar_corpus
  data_files:
  - split: train
    path: data/edgar_corpus/train-*.parquet
  - split: test
    path: data/edgar_corpus/test-*.parquet
  - split: validate
    path: data/edgar_corpus/validate-*.parquet
- config_name: eli5
  data_files:
  - split: train
    path: data/eli5/train-*.parquet
  - split: test
    path: data/eli5/test-*.parquet
  - split: validation
    path: data/eli5/validation-*.parquet
- config_name: elife_plos_lay_summaries
  data_files:
  - split: train
    path: data/elife_plos_lay_summaries/train-*.parquet
  - split: test
    path: data/elife_plos_lay_summaries/test-*.parquet
  - split: validation
    path: data/elife_plos_lay_summaries/validation-*.parquet
- config_name: europarl
  data_files:
  - split: full
    path: data/europarl/full-*.parquet
- config_name: govreport
  data_files:
  - split: train
    path: data/govreport/train-*.parquet
  - split: test
    path: data/govreport/test-*.parquet
  - split: valid
    path: data/govreport/valid-*.parquet
- config_name: hh_rlhf_harmless_base
  data_files:
  - split: train
    path: data/hh_rlhf_harmless_base/train-*.parquet
  - split: test
    path: data/hh_rlhf_harmless_base/test-*.parquet
- config_name: hh_rlhf_helpful_base
  data_files:
  - split: train
    path: data/hh_rlhf_helpful_base/train-*.parquet
  - split: test
    path: data/hh_rlhf_helpful_base/test-*.parquet
- config_name: hh_rlhf_helpful_online
  data_files:
  - split: train
    path: data/hh_rlhf_helpful_online/train-*.parquet
  - split: test
    path: data/hh_rlhf_helpful_online/test-*.parquet
- config_name: hh_rlhf_helpful_rejection_sampled
  data_files:
  - split: train
    path: data/hh_rlhf_helpful_rejection_sampled/train-*.parquet
  - split: test
    path: data/hh_rlhf_helpful_rejection_sampled/test-*.parquet
- config_name: hh_rlhf_red_team_attempts
  data_files:
  - split: full
    path: data/hh_rlhf_red_team_attempts/full-*.parquet
- config_name: ifbench
  data_files:
  - split: test
    path: data/ifbench/test-*.parquet
- config_name: issuebench
  data_files:
  - split: prompts_full
    path: data/issuebench/prompts_full-*.parquet
- config_name: issuebench_issues
  data_files:
  - split: issues
    path: data/issuebench_issues/issues-*.parquet
- config_name: issuebench_templates
  data_files:
  - split: templates_full
    path: data/issuebench_templates/templates_full-*.parquet
- config_name: iterater_full_doc
  data_files:
  - split: train
    path: data/iterater_full_doc/train-*.parquet
  - split: dev
    path: data/iterater_full_doc/dev-*.parquet
  - split: test
    path: data/iterater_full_doc/test-*.parquet
- config_name: iterater_human_doc
  data_files:
  - split: train
    path: data/iterater_human_doc/train-*.parquet
  - split: dev
    path: data/iterater_human_doc/dev-*.parquet
  - split: test
    path: data/iterater_human_doc/test-*.parquet
- config_name: kilt
  data_files:
  - split: full
    path: data/kilt/full-*.parquet
- config_name: kiwi
  data_files:
  - split: full
    path: data/kiwi/full-*.parquet
- config_name: klicke
  data_files:
  - split: full
    path: data/klicke/full-*.parquet
- config_name: klicke_keystrokes
  data_files:
  - split: full
    path: data/klicke_keystrokes/full-*.parquet
- config_name: klicke_typing_tests
  data_files:
  - split: full
    path: data/klicke_typing_tests/full-*.parquet
- config_name: klicke_vocabulary_test
  data_files:
  - split: full
    path: data/klicke_vocabulary_test/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: arxiv_set
    path: data/l1_signals/arxiv_set-*.parquet
- config_name: longbench_write
  data_files:
  - split: full
    path: data/longbench_write/full-*.parquet
- config_name: longbench_write_ruler
  data_files:
  - split: full
    path: data/longbench_write_ruler/full-*.parquet
- config_name: masc
  data_files:
  - split: full
    path: data/masc/full-*.parquet
- config_name: masc_oanc
  data_files:
  - split: full
    path: data/masc_oanc/full-*.parquet
- config_name: medquad
  data_files:
  - split: full
    path: data/medquad/full-*.parquet
- config_name: methods
  data_files:
  - split: full
    path: data/methods/full-*.parquet
- config_name: million_authors
  data_files:
  - split: full
    path: data/million_authors/full-*.parquet
- config_name: million_authors_av
  data_files:
  - split: train
    path: data/million_authors_av/train-*.parquet
  - split: dev
    path: data/million_authors_av/dev-*.parquet
  - split: test
    path: data/million_authors_av/test-*.parquet
- config_name: mt_bench
  data_files:
  - split: train
    path: data/mt_bench/train-*.parquet
- config_name: muldiloc
  data_files:
  - split: full
    path: data/muldiloc/full-*.parquet
- config_name: multieurlex
  data_files:
  - split: train
    path: data/multieurlex/train-*.parquet
  - split: test
    path: data/multieurlex/test-*.parquet
  - split: validation
    path: data/multieurlex/validation-*.parquet
- config_name: newswire
  data_files:
  - split: full
    path: data/newswire/full-*.parquet
- config_name: nirvana
  data_files:
  - split: full
    path: data/nirvana/full-*.parquet
- config_name: parlamint
  data_files:
  - split: full
    path: data/parlamint/full-*.parquet
- config_name: parliament_undisclosed_llm
  data_files:
  - split: train
    path: data/parliament_undisclosed_llm/train-*.parquet
  - split: test
    path: data/parliament_undisclosed_llm/test-*.parquet
  - split: validation
    path: data/parliament_undisclosed_llm/validation-*.parquet
- config_name: peerread
  data_files:
  - split: train
    path: data/peerread/train-*.parquet
  - split: dev
    path: data/peerread/dev-*.parquet
  - split: test
    path: data/peerread/test-*.parquet
- config_name: peerread_papers
  data_files:
  - split: train
    path: data/peerread_papers/train-*.parquet
  - split: dev
    path: data/peerread_papers/dev-*.parquet
  - split: test
    path: data/peerread_papers/test-*.parquet
- config_name: pg19
  data_files:
  - split: train
    path: data/pg19/train-*.parquet
  - split: test
    path: data/pg19/test-*.parquet
  - split: validation
    path: data/pg19/validation-*.parquet
- config_name: prism
  data_files:
  - split: full
    path: data/prism/full-*.parquet
- config_name: project_gutenberg
  data_files:
  - split: full
    path: data/project_gutenberg/full-*.parquet
- config_name: promptevals
  data_files:
  - split: train
    path: data/promptevals/train-*.parquet
  - split: test
    path: data/promptevals/test-*.parquet
  - split: validation
    path: data/promptevals/validation-*.parquet
- config_name: rcv1_reuter_50_50
  data_files:
  - split: train
    path: data/rcv1_reuter_50_50/train-*.parquet
  - split: test
    path: data/rcv1_reuter_50_50/test-*.parquet
- config_name: realguardrails_distractors
  data_files:
  - split: full
    path: data/realguardrails_distractors/full-*.parquet
- config_name: realguardrails_handwritten
  data_files:
  - split: full
    path: data/realguardrails_handwritten/full-*.parquet
- config_name: realguardrails_prompts
  data_files:
  - split: full
    path: data/realguardrails_prompts/full-*.parquet
- config_name: reddit_tifu
  data_files:
  - split: full
    path: data/reddit_tifu/full-*.parquet
- config_name: slopbench
  data_files:
  - split: heldout
    path: data/slopbench/heldout-*.parquet
  - split: public
    path: data/slopbench/public-*.parquet
- config_name: stack_exchange
  data_files:
  - split: full
    path: data/stack_exchange/full-*.parquet
- config_name: supreme_court_oral_arguments
  data_files:
  - split: full
    path: data/supreme_court_oral_arguments/full-*.parquet
- config_name: un_general_debates
  data_files:
  - split: full
    path: data/un_general_debates/full-*.parquet
- config_name: un_parallel
  data_files:
  - split: full
    path: data/un_parallel/full-*.parquet
- config_name: un_parallel_testsets
  data_files:
  - split: dev
    path: data/un_parallel_testsets/dev-*.parquet
  - split: test
    path: data/un_parallel_testsets/test-*.parquet
- config_name: webis_tldr_17
  data_files:
  - split: full
    path: data/webis_tldr_17/full-*.parquet
- config_name: wikiconv
  data_files:
  - split: full
    path: data/wikiconv/full-*.parquet
- config_name: wikipedia
  data_files:
  - split: full
    path: data/wikipedia/full-*.parquet
- config_name: wildchat
  data_files:
  - split: train
    path: data/wildchat/train-*.parquet
---

# Anagram Corpus

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 allows commercial use.** Non-commercial material is in the companion repository [CoderBak/anagram-corpus-nc](https://huggingface.co/datasets/CoderBak/anagram-corpus-nc); load both for the full corpus (research use).

This version holds **820,570,578 rows** in **85 subsets** (one per source part). It is being built source by source; more subsets will follow.

## Loading

```python
from datasets import load_dataset
ds = load_dataset("CoderBak/anagram-corpus", "ai_text_not_mine")   # 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 |
|---|---|---|---|---|---|
| `ai_text_not_mine` | documents | 95 | test 95 | CC BY-SA 4.0 | [AITDNA ('Your AI Text is not Mine'): human-only texts and their edit logs](https://arxiv.org/abs/2606.04906) |
| `ai_text_not_mine_keystrokes` | logs | 95 | test 95 | CC BY-SA 4.0 | [AITDNA ('Your AI Text is not Mine'): human-only texts and their edit logs](https://arxiv.org/abs/2606.04906) |
| `american_stories` | documents | 370,313,913 | full 370,313,913 | CC BY 4.0 | [American Stories: A Large-Scale Structured Text Dataset of Historical U.S. Newspapers](https://arxiv.org/abs/2308.12477) |
| `arxiv` | documents | 3,205,432 | full 3,205,432 | CC0 1.0 | [arXiv metadata snapshot (titles and abstracts)](https://arxiv.org/abs/1905.00075) |
| `asap_2` | documents | 24,728 | test 7,421, train 17,307 | CC BY 4.0 | [ASAP 2.0 (Automated Student Assessment Prize 2.0 corpus)](https://zenodo.org/records/14781349) |
| `caselaw_access_project` | documents | 6,920,587 | full 6,920,587 | CC0 1.0 | [Caselaw Access Project (Harvard Law School Library)](https://case.law/) |
| `casimir` | documents | 54,675 | test 5,579, train 43,615, validation 5,481 | MIT | [CASIMIR: A Corpus of Scientific Articles Enhanced with Multiple Author-Integrated Revisions](https://aclanthology.org/2024.lrec-main.257/) |
| `casimir_reviews` | documents | 185,991 | full 185,991 | MIT | [CASIMIR: A Corpus of Scientific Articles Enhanced with Multiple Author-Integrated Revisions](https://aclanthology.org/2024.lrec-main.257/) |
| `chatbot_arena_140k` | prompts | 135,634 | train 135,634 | CC BY 4.0 | [Chatbot Arena human preference votes (arena-human-preference-140k and -55k)](https://proceedings.mlr.press/v235/chiang24b.html) |
| `chatbot_arena_55k` | prompts | 57,477 | train 57,477 | Apache 2.0 | [Chatbot Arena human preference votes (arena-human-preference-140k and -55k)](https://proceedings.mlr.press/v235/chiang24b.html) |
| `chatgpt_detector_bias` | documents | 394 | full 394 | MIT | [GPT Detectors Are Biased Against Non-Native English Writers (human-written data)](https://arxiv.org/abs/2304.02819) |
| `civil_comments` | documents | 1,999,514 | test 97,320, train 1,804,874, validation 97,320 | CC0 1.0 | [Civil Comments (Jigsaw Unintended Bias in Toxicity Classification)](https://arxiv.org/abs/1903.04561) |
| `cmu_movie_summaries` | documents | 42,306 | full 42,306 | CC BY-SA 3.0 | [CMU Movie Summary Corpus](https://aclanthology.org/P13-1035/) |
| `co_storm_wildseek` | prompts | 100 | full 100 | CC BY-SA 4.0 | [WildSeek (Co-STORM: Into the Unknown Unknowns)](https://arxiv.org/abs/2408.15232) |
| `common_pile_foodista` | documents | 72,034 | train 72,034 | CC BY 3.0 | [Common Pile v0.1 (raw): Ubuntu IRC, Foodista, Regulations.gov, USGPO, GitHub Archive](https://arxiv.org/abs/2506.05209) |
| `common_pile_github_archive` | documents | 30,318,756 | train 30,318,756 | MIT, Apache 2.0, BSD 3-Clause, CC BY 4.0, BSD 2-Clause, CC0 1.0, The Unlicense, ISC, Boost Software License 1.0, zlib License, WTFPL, MIT No Attribution, BSD Zero Clause, Artistic License 2.0, Vim License, Universal Permissive License 1.0, Educational Community License 2.0, BSD 3-Clause Clear, Microsoft Public License, Academic Free License 3.0, LaTeX Project Public License 1.3c, PostgreSQL License, University of Illinois/NCSA Open Source License, BSD 4-Clause, Mulan Permissive Software License 2.0 | [Common Pile v0.1 (raw): Ubuntu IRC, Foodista, Regulations.gov, USGPO, GitHub Archive](https://arxiv.org/abs/2506.05209) |
| `common_pile_regulations` | documents | 222,980 | train 222,980 | Public domain (U.S. government work or expired copyright) | [Common Pile v0.1 (raw): Ubuntu IRC, Foodista, Regulations.gov, USGPO, GitHub Archive](https://arxiv.org/abs/2506.05209) |
| `common_pile_ubuntu_irc` | documents | 329,113 | train 329,113 | Public domain (U.S. government work or expired copyright) | [Common Pile v0.1 (raw): Ubuntu IRC, Foodista, Regulations.gov, USGPO, GitHub Archive](https://arxiv.org/abs/2506.05209) |
| `common_pile_usgpo` | documents | 2,732,677 | train 2,732,677 | Public domain (U.S. government work or expired copyright) | [Common Pile v0.1 (raw): Ubuntu IRC, Foodista, Regulations.gov, USGPO, GitHub Archive](https://arxiv.org/abs/2506.05209) |
| `congressional_hearings` | documents | 35,696 | full 35,696 | Public domain (U.S. government work or expired copyright) | [Congressional hearings (govinfo CHRG)](https://www.govinfo.gov/app/collection/chrg) |
| `congressional_record` | documents | 910,669 | full 910,669 | Public domain (U.S. government work or expired copyright) | [Congressional Record (daily edition, govinfo CREC)](https://www.govinfo.gov/app/collection/crec) |
| `courtlistener` | documents | 10,901,752 | full 10,901,752 | Public domain (U.S. government work or expired copyright) | [CourtListener bulk opinions (Free Law Project), snapshot 2026-09-30](https://www.courtlistener.com/help/api/bulk-data/) |
| `cuad` | documents | 510 | test 102, train 408 | CC BY 4.0 | [CUAD: Contract Understanding Atticus Dataset v1](https://arxiv.org/abs/2103.06268) |
| `djinni` | documents | 141,897 | train 141,897 | MIT | [Djinni Recruitment Dataset (English job descriptions)](https://aclanthology.org/2024.unlp-1.2/) |
| `ebg_stylometry` | documents | 802 | full 802 | BSD 3-Clause | [Extended Brennan-Greenstadt and Brennan-Greenstadt adversarial stylometry corpora](https://doi.org/10.1145/2382448.2382450) |
| `ebg_stylometry_bg` | documents | 102 | full 102 | BSD 3-Clause | [Extended Brennan-Greenstadt and Brennan-Greenstadt adversarial stylometry corpora](https://doi.org/10.1145/2382448.2382450) |
| `edgar_corpus` | documents | 218,768 | test 21,890, train 174,986, validate 21,892 | CC BY 4.0 | [EDGAR-CORPUS: Billions of Tokens Make The World Go Round](https://aclanthology.org/2021.econlp-1.2/) |
| `eli5` | prompts | 274,736 | test 600, train 272,629, validation 1,507 | MIT | [ELI5: Long Form Question Answering (KILT release)](https://aclanthology.org/P19-1346/) |
| `elife_plos_lay_summaries` | documents | 32,353 | test 1,617, train 29,119, validation 1,617 | CC BY 4.0 | [PLOS and eLife lay summarisation corpora (Making Science Simple)](https://aclanthology.org/2022.emnlp-main.724/) |
| `europarl` | documents | 2,787,578 | full 2,787,578 | Free reuse with acknowledgement of the source (official publications) | [Europarl v7: European Parliament Proceedings Parallel Corpus 1996-2011 (source release)](https://aclanthology.org/2005.mtsummit-papers.11/) |
| `govreport` | documents | 19,466 | test 973, train 17,519, valid 974 | CC BY 4.0 | [GovReport](https://aclanthology.org/2021.naacl-main.112/) |
| `hh_rlhf_harmless_base` | prompts | 44,849 | test 2,312, train 42,537 | MIT | [Anthropic HH-RLHF (helpfulness and harmlessness preference data, red-team attempts)](https://arxiv.org/abs/2204.05862) |
| `hh_rlhf_helpful_base` | prompts | 46,189 | test 2,354, train 43,835 | MIT | [Anthropic HH-RLHF (helpfulness and harmlessness preference data, red-team attempts)](https://arxiv.org/abs/2204.05862) |
| `hh_rlhf_helpful_online` | prompts | 23,144 | test 1,137, train 22,007 | MIT | [Anthropic HH-RLHF (helpfulness and harmlessness preference data, red-team attempts)](https://arxiv.org/abs/2204.05862) |
| `hh_rlhf_helpful_rejection_sampled` | prompts | 55,170 | test 2,749, train 52,421 | MIT | [Anthropic HH-RLHF (helpfulness and harmlessness preference data, red-team attempts)](https://arxiv.org/abs/2204.05862) |
| `hh_rlhf_red_team_attempts` | prompts | 38,961 | full 38,961 | MIT | [Anthropic HH-RLHF (helpfulness and harmlessness preference data, red-team attempts)](https://arxiv.org/abs/2204.05862) |
| `ifbench` | prompts | 300 | test 300 | ODC Attribution 1.0 | [IFBench: Generalizing Verifiable Instruction Following](https://arxiv.org/abs/2507.02833) |
| `issuebench` | prompts | 2,490,576 | prompts_full 2,490,576 | CC BY 4.0 | [IssueBench](https://arxiv.org/abs/2502.08395) |
| `issuebench_issues` | prompts | 396 | issues 396 | CC BY 4.0 | [IssueBench](https://arxiv.org/abs/2502.08395) |
| `issuebench_templates` | prompts | 3,916 | templates_full 3,916 | CC BY 4.0 | [IssueBench](https://arxiv.org/abs/2502.08395) |
| `iterater_full_doc` | documents | 63,262 | dev 1,712, test 1,854, train 59,696 | CC BY-SA 3.0, Apache 2.0, CC BY 2.5 | [IteraTeR (document level: IteraTeR-FULL and IteraTeR-HUMAN)](https://aclanthology.org/2022.acl-long.250/) |
| `iterater_human_doc` | documents | 1,118 | dev 54, test 102, train 962 | CC BY 2.5, CC BY-SA 3.0, Apache 2.0 | [IteraTeR (document level: IteraTeR-FULL and IteraTeR-HUMAN)](https://aclanthology.org/2022.acl-long.250/) |
| `kilt` | documents | 5,889,353 | full 5,889,353 | CC BY-SA 3.0 | [KILT knowledge source (English Wikipedia, 2019-08-01)](https://aclanthology.org/2021.naacl-main.200/) |
| `kiwi` | prompts | 234 | full 234 | CC BY-SA 4.0 | [KIWI: A Dataset of Knowledge-Intensive Writing Instructions for Answering Research Questions](https://aclanthology.org/2024.findings-acl.770/) |
| `klicke` | documents | 4,992 | full 4,992 | MIT | [KLiCKe: Keystroke Logging in Compositions for Knowledge Evaluation](https://www.jowr.org/jowr/article/view/1556) |
| `klicke_keystrokes` | logs | 9,984 | full 9,984 | MIT | [KLiCKe: Keystroke Logging in Compositions for Knowledge Evaluation](https://www.jowr.org/jowr/article/view/1556) |
| `klicke_typing_tests` | logs | 69,840 | full 69,840 | MIT | [KLiCKe: Keystroke Logging in Compositions for Knowledge Evaluation](https://www.jowr.org/jowr/article/view/1556) |
| `klicke_vocabulary_test` | logs | 4,971 | full 4,971 | MIT | [KLiCKe: Keystroke Logging in Compositions for Knowledge Evaluation](https://www.jowr.org/jowr/article/view/1556) |
| `l1_signals` | documents | 2,357 | acl_postllm 369, acl_prellm 388, arxiv_set 1,600 | CC0 1.0, CC BY 4.0 | [Can We Still Trace L1 Signals? (academic abstracts labelled with the authors' native language)](https://arxiv.org/abs/2604.08568) |
| `longbench_write` | prompts | 120 | full 120 | Apache 2.0 | [LongBench-Write (LongWriter)](https://arxiv.org/abs/2408.07055) |
| `longbench_write_ruler` | prompts | 48 | full 48 | Apache 2.0 | [LongBench-Write (LongWriter)](https://arxiv.org/abs/2408.07055) |
| `masc` | documents | 392 | full 392 | CC BY 3.0 | [MASC 3.0.0 and the Open American National Corpus (OANC)](https://aclanthology.org/P10-2013/) |
| `masc_oanc` | documents | 8,824 | full 8,824 | CC BY 3.0 | [MASC 3.0.0 and the Open American National Corpus (OANC)](https://aclanthology.org/P10-2013/) |
| `medquad` | documents | 16,423 | full 16,423 | CC BY 4.0 | [MedQuAD: Medical Question Answering Dataset](https://doi.org/10.1186/s12859-019-3119-4) |
| `methods` | methods | 8,670 | full 8,670 | Apache 2.0, MIT, CC BY 4.0, Not copyrightable (settings and parameters), BSD 3-Clause, CC BY-SA 4.0, CC BY 3.0, GPL 3.0 | anagram-bench generation methods |
| `million_authors` | documents | 60,083,121 | full 60,083,121 | CC BY-SA 4.0 | [The Million Authors Corpus (MAC)](https://aclanthology.org/2025.findings-acl.1335/) |
| `million_authors_av` | documents | 2,700,034 | dev 270,020, test 540,054, train 1,889,960 | CC BY-SA 4.0 | [The Million Authors Corpus (MAC)](https://aclanthology.org/2025.findings-acl.1335/) |
| `mt_bench` | prompts | 33,000 | train 33,000 | CC BY 4.0 | [Chatbot Arena Conversations (MT-Bench and Chatbot Arena paper)](https://arxiv.org/abs/2306.05685) |
| `muldiloc` | documents | 1,417 | full 1,417 | MIT | [MulDiLoC-HAP: Multidialectal, Longitudinal Corpus of Human-AI Hybrid Production (human-written essays)](https://aclanthology.org/2026.lrec-1.882) |
| `multieurlex` | documents | 1,044,739 | test 115,000, train 817,239, validation 112,500 | CC BY-SA 4.0 | [MultiEURLEX: multilingual EU legislation with EUROVOC labels](https://aclanthology.org/2021.emnlp-main.559) |
| `newswire` | documents | 2,719,607 | full 2,719,607 | CC BY 4.0 | [Newswire: A Large-Scale Structured Database of a Century of Historical News](https://arxiv.org/abs/2406.09490) |
| `nirvana` | prompts | 71 | full 71 | CC BY 4.0 | [NIRVANA: A Comprehensive Dataset for Reproducing How Students Use Generative AI for Essay Writing](https://arxiv.org/abs/2604.07344) |
| `parlamint` | documents | 8,073,406 | full 8,073,406 | CC BY 4.0 | [ParlaMint 5.0: comparable corpora of parliamentary debates (29 parliaments)](https://doi.org/10.1007/s10579-021-09574-0) |
| `parliament_undisclosed_llm` | documents | 17,410 | test 1,742, train 13,927, validation 1,741 | CC BY 4.0 | [Detecting undisclosed LLM-generated content in parliamentary texts (data release)](https://arxiv.org/abs/2606.14209) |
| `peerread` | documents | 314 | dev 13, test 20, train 281 | CC BY 4.0 | [PeerRead: A Dataset of Peer Reviews (ACL 2017 and CoNLL 2016 sections)](https://aclanthology.org/N18-1149/) |
| `peerread_papers` | documents | 159 | dev 8, test 9, train 142 | CC BY 4.0 | [PeerRead: A Dataset of Peer Reviews (ACL 2017 and CoNLL 2016 sections)](https://aclanthology.org/N18-1149/) |
| `pg19` | documents | 28,752 | test 100, train 28,602, validation 50 | Apache 2.0 | [PG-19 Language Modelling Benchmark](https://arxiv.org/abs/1911.05507) |
| `prism` | prompts | 8,011 | full 8,011 | CC BY 4.0 | [The PRISM Alignment Dataset](https://arxiv.org/abs/2404.16019) |
| `project_gutenberg` | documents | 77,588 | full 77,588 | Public domain (U.S. government work or expired copyright) | [Project Gutenberg (bulk plain-text feed)](https://www.gutenberg.org/) |
| `promptevals` | prompts | 2,088 | test 419, train 1,252, validation 417 | MIT | [PromptEvals: A Dataset of Assertions and Guardrails for Custom Production LLM Pipelines](https://aclanthology.org/2025.naacl-long.213/) |
| `rcv1_reuter_50_50` | documents | 5,000 | test 2,500, train 2,500 | CC BY 4.0 | [Reuter_50_50 (RCV1 authorship subset)](https://jmlr.org/papers/v5/lewis04a.html) |
| `realguardrails_distractors` | prompts | 504 | full 504 | CC BY 4.0, CC BY-SA 4.0 | [RealGuardrails / SystemCheck: A Closer Look at System Prompt Robustness](https://arxiv.org/abs/2502.12197) |
| `realguardrails_handwritten` | prompts | 239 | full 239 | MIT | [RealGuardrails / SystemCheck: A Closer Look at System Prompt Robustness](https://arxiv.org/abs/2502.12197) |
| `realguardrails_prompts` | prompts | 3,082 | full 3,082 | MIT | [RealGuardrails / SystemCheck: A Closer Look at System Prompt Robustness](https://arxiv.org/abs/2502.12197) |
| `reddit_tifu` | documents | 79,750 | full 79,750 | MIT | [Reddit TIFU](https://aclanthology.org/N19-1260/) |
| `slopbench` | prompts | 112 | heldout 38, public 74 | MIT | [SlopBench: How Well Can We Rank Language Models by Slop?](https://arxiv.org/abs/2609.33905) |
| `stack_exchange` | documents | 82,921,221 | full 82,921,221 | CC BY-SA 3.0, CC BY-SA 4.0, CC BY-SA 2.5 | [Stack Exchange Data Dump (2024-04-02)](https://archive.org/details/stackexchange) |
| `supreme_court_oral_arguments` | documents | 1,809 | full 1,809 | Public domain (U.S. government work or expired copyright) | [U.S. Supreme Court oral argument transcripts (OT2000-OT2026)](https://www.supremecourt.gov/oral_arguments/availabilityoforalargumenttranscripts.aspx) |
| `un_general_debates` | documents | 8,093 | full 8,093 | CC0 1.0 | [UN General Debate Corpus (UNGDC) 1970-2018](https://doi.org/10.1177/2053168017712821) |
| `un_parallel` | documents | 799,173 | full 799,173 | Public domain (U.S. government work or expired copyright) | [The United Nations Parallel Corpus v1.0](https://aclanthology.org/L16-1561/) |
| `un_parallel_testsets` | documents | 48,000 | dev 24,000, test 24,000 | Public domain (U.S. government work or expired copyright) | [The United Nations Parallel Corpus v1.0](https://aclanthology.org/L16-1561/) |
| `webis_tldr_17` | documents | 3,848,330 | full 3,848,330 | CC BY 4.0 | [Webis-TLDR-17](https://aclanthology.org/W17-4508/) |
| `wikiconv` | documents | 210,018,590 | full 210,018,590 | CC BY-SA 3.0 | [WikiConv (English)](https://aclanthology.org/D18-1305/) |
| `wikipedia` | documents | 6,487,681 | full 6,487,681 | CC BY-SA 3.0 | [English Wikipedia (Wikimedia dump enwiki-20220501)](https://dumps.wikimedia.org/) |
| `wildchat` | prompts | 834,358 | train 834,358 | ODC Attribution 1.0 | [WildChat-1M](https://arxiv.org/abs/2405.01470) |

## Columns

Every row has: `id` (<source>/<source_record_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

### `ai_text_not_mine`

AITDNA ('Your AI Text is not Mine'): human-only texts and their edit logs. Dycke, Sakharova, Daheim & Gurevych (arXiv 2026).

The release holds 362 texts from user studies (one split, 'test'), each in nine configs that share rows and metadata: 'original' (the editor log of the text: every insert, delete and retain event with offset, text, seconds since the first edit, and user), 'document' (the final text with its authorship label) and seven other label views (sentence, token, span, boundary, intent, content, membership). Each participant wrote three texts with an LLM assistant and one with LLM support turned off (metadata human_only = true). Included: the 95 human_only texts (subset ai_text_not_mine, documents) and the edit log of each of them (subset ai_text_not_mine_keystrokes, logs, one row per text). Left out: the 267 texts written with LLM support on (human_only = false; co-written with gpt-5.2, deepseek-chat, qwen2.5:7b, llama4:scout or gemini-3-flash-preview, including the 10 peer reviews) and their logs, because they contain or may contain AI-generated text. The seven label views add nothing for human-only texts (every label is 'User') and are not used. Rows keep the release order; `source_record_id` = 'aitdna_test_<0-based row index>' (the release has no ids). Documents: `text` = the 'document' config's text segments joined (one segment, author 'User', for every included text; it equals the replay of the log (up to trailing newlines) for 89 of the 95 and differs by 1-2 characters in 6); `genre` from metadata.task; `author_id` = metadata.author (participant pseudonym; links the four texts of a session); `prompt` empty (the participant guidelines and topic lists are only in the paper's figures, and which of two offered topics a writer chose is not released); `meta` = the release's metadata (author, human_only, model, temperature, setting, task). `date` = 2026, an upper bound (paper v1 2026-06-03; per-text dates are not released): the sessions' co-writing used models released in 2024-2025 (up to GPT-5.2 and Gemini 3 Flash, December 2025), so every text was written after 2022-11-30. Logs: `data` = the text's 'original' config event list as a JSON array whose elements are the released event objects in release order (format 'json'); `parent_id` = the document row id; `meta` = event counts by operationType and user, and the release metadata. The logs are editor change events batched about every 2 s, not single keystrokes; all events of the included logs are by 'User' (no LLM request or response events).

*Licence:* Stated inconsistently: the HF card's YAML says cc-by-sa-4.0, the card body says 'License: gpl-2.0' (the code repo is Apache-2.0, the paper CC BY 4.0). The corpus has no GPL-2.0 licence id; rows carry CC BY-SA 4.0, the licence in the card metadata. Both readings allow commercial use and require attribution and share-alike (copyleft) for adaptations, so the repo is the same (anagram-corpus) either way. The texts are by anonymised study participants (pseudonymous author names).

### `ai_text_not_mine_keystrokes`

AITDNA ('Your AI Text is not Mine'): human-only texts and their edit logs. Dycke, Sakharova, Daheim & Gurevych (arXiv 2026).

The release holds 362 texts from user studies (one split, 'test'), each in nine configs that share rows and metadata: 'original' (the editor log of the text: every insert, delete and retain event with offset, text, seconds since the first edit, and user), 'document' (the final text with its authorship label) and seven other label views (sentence, token, span, boundary, intent, content, membership). Each participant wrote three texts with an LLM assistant and one with LLM support turned off (metadata human_only = true). Included: the 95 human_only texts (subset ai_text_not_mine, documents) and the edit log of each of them (subset ai_text_not_mine_keystrokes, logs, one row per text). Left out: the 267 texts written with LLM support on (human_only = false; co-written with gpt-5.2, deepseek-chat, qwen2.5:7b, llama4:scout or gemini-3-flash-preview, including the 10 peer reviews) and their logs, because they contain or may contain AI-generated text. The seven label views add nothing for human-only texts (every label is 'User') and are not used. Rows keep the release order; `source_record_id` = 'aitdna_test_<0-based row index>' (the release has no ids). Documents: `text` = the 'document' config's text segments joined (one segment, author 'User', for every included text; it equals the replay of the log (up to trailing newlines) for 89 of the 95 and differs by 1-2 characters in 6); `genre` from metadata.task; `author_id` = metadata.author (participant pseudonym; links the four texts of a session); `prompt` empty (the participant guidelines and topic lists are only in the paper's figures, and which of two offered topics a writer chose is not released); `meta` = the release's metadata (author, human_only, model, temperature, setting, task). `date` = 2026, an upper bound (paper v1 2026-06-03; per-text dates are not released): the sessions' co-writing used models released in 2024-2025 (up to GPT-5.2 and Gemini 3 Flash, December 2025), so every text was written after 2022-11-30. Logs: `data` = the text's 'original' config event list as a JSON array whose elements are the released event objects in release order (format 'json'); `parent_id` = the document row id; `meta` = event counts by operationType and user, and the release metadata. The logs are editor change events batched about every 2 s, not single keystrokes; all events of the included logs are by 'User' (no LLM request or response events).

*Licence:* Stated inconsistently: the HF card's YAML says cc-by-sa-4.0, the card body says 'License: gpl-2.0' (the code repo is Apache-2.0, the paper CC BY 4.0). The corpus has no GPL-2.0 licence id; rows carry CC BY-SA 4.0, the licence in the card metadata. Both readings allow commercial use and require attribution and share-alike (copyleft) for adaptations, so the repo is the same (anagram-corpus) either way. The texts are by anonymised study participants (pseudonymous author names).

### `american_stories`

American Stories: A Large-Scale Structured Text Dataset of Historical U.S. Newspapers. Dell, Carlson, Bryan, Silcock, Arora, Shen, D'Amico-Wong, Le, Querubin & Heldring (NeurIPS 2023 Datasets and Benchmarks).

Every associated article ('full articles' entry) of every scan JSON in all 180 year archives (1770-1964): 370,313,913 articles from 16,887,460 scanned pages, in archive order; nothing sampled or filtered. This is the unit of the release's article-level output (official loader AmericanStories.py, config all_years); `source_record_id` is the loader's article_id, '<full_article_id>_<scan file name without .json>' (unique across the release). Every full article is built from exactly one 'article' content region plus the 'headline' and 'author' regions the release associated with it: `text` = the release's `article` field (the article region's OCR text as released, OCR errors and hard line breaks kept), `title` = `headline` (the headline regions' OCR joined by blank lines; None when empty), meta.byline = `byline` (the author regions joined). `date` = the issue date from the scan file name (the loader's `date`, YYYY-MM-DD; 1770-08-23 .. 1964-12-31, present for every article); `parent_id` = the scan id (file name without .json; one scan = one newspaper page); `source_url` = the Chronicling America page image (scan.raw_data_loc, else scan_url; None for the 1,164,263 articles whose scan record has neither); `genre` = 'newspaper article (historical, OCR)'; no author ids. `meta`: year_file and scan_file (archive and member name); the loader's fields article_id, newspaper_name (lccn.title), edition and page (parsed from the file name as the loader does); the full-article fields as released (full_article_id, id, byline, object_ids, bbox, bbox_list); objects = the release's record of each content region of the article (id, class, legibility Legible / Questionable / Illegible / NA, reading_order_id, full_article_id, bbox); and scan_record = the scan-level metadata as released (lccn = the newspaper's Library of Congress record: title, lccn, state, issn, start/end year, language(s), geonames/dbpedia ids, ...; edition; scan = page size and Chronicling America URLs; page_number; id; scan_url and scan_ocr where present). The regions' raw_text is not repeated in objects because the article fields restate it exactly (`article` = the article region's raw_text, `headline` and `byline` = the headline and author regions' raw_text joined by blank lines; checked for every article, true for all). Included although the official loader skips them: the articles of the 161,322 scans whose JSON has no `lccn` record (scan_record.lccn = null; the newspaper's LCCN is still in the file name). Left out: content regions that belong to no full article, which the release's article-level output does not contain either: 18,536,873 article regions, all with empty raw_text (regions not OCR'd, mostly those the legibility classifier marked Illegible or NA), 7,718,136 headline and 230,677 author regions with text but no associated article (stand-alone headlines and bylines), and 7,725,190 empty headline/author regions. No scan file failed to parse and no article has empty text, so the builder dropped nothing. (Code path not taken: an article_id repeating within a year archive would get the suffix '~dup<k>' and meta.duplicate_article_id = true.)

*Licence:* CC BY 4.0 (dataset card 'license: cc-by-4.0'; LICENSE file = CC BY 4.0 legal code; README: 'The dataset has a CC-BY 4.0 license'). Attribution required. The scans are public-domain Chronicling America pages (Library of Congress); the paper notes that some syndicated content (e.g. comics, fiction) in the newspapers may still be under copyright. Historical newspapers contain period offensive language and personal details of people named in the news.

### `arxiv`

arXiv metadata snapshot (titles and abstracts). Cornell University, arXiv Dataset (Kaggle Cornell-University/arxiv); Clement, Bierbaum, O'Keeffe & Alemi, On the Use of ArXiv as a Dataset (2019).

Every record of the snapshot: 3,205,432 papers, one row each, in file order. Nothing is filtered. `text` = the abstract with arXiv's line wrapping undone by arXiv's own rule ('Carriage returns will be stripped unless they are followed by leading white spaces', info.arxiv.org/help/prep): a line break followed by an indented or empty line is kept as a line break (blank lines kept, runs of them cut to one), every other line break becomes a space, and the indentation and the line ends' spaces are dropped; otherwise the abstract is as released (LaTeX markup such as $...$ kept). `title` = the title with its wrapping whitespace collapsed to single spaces. The snapshot holds each paper's current metadata, so title and abstract are those of the latest version. `date` = the creation date of that latest version (UTC), so pre_chatgpt is true only if the text as released was written before 2022-11-30; the first version's date and every other version date are in meta.versions. `meta` = the record's other fields as released: submitter, authors, authors_parsed, comments, journal-ref, doi, report-no, categories, license (the licence the authors chose for the paper itself, not for this metadata), versions (every version with its created date), update_date. `source_record_id` = arXiv id; 80 ids occur more than once (old math-ph ids, different records with different titles and abstracts): all occurrences are kept, the 2nd and later get '#2', '#3', ... `source_url` = https://arxiv.org/abs/<id>. Full texts are not part of this release.

*Licence:* CC0 1.0: arXiv dedicates all metadata, including titles and abstracts, to the public domain (info.arxiv.org/help/license: 'A Creative Commons CC0 1.0 Universal Public Domain Dedication will apply to all metadata'); Kaggle lists the dataset as 'CC0: Public Domain'. The papers' own licences (meta.license) cover the full texts, which are not included.

### `asap_2`

ASAP 2.0 (Automated Student Assessment Prize 2.0 corpus). Crossley et al., 'A large-scale corpus for assessing source-based writing quality: ASAP 2.0' (Assessing Writing 65, 2025; preprint Zenodo 14781349).

Every essay of both official files: train = every row of ASAP_2_Final_github_train.csv in the training zip (17,307 essays), test = every row of ASAP_2_Final_github_test.csv in the password-protected test zip (7,421; password from the README); 24,728 in all (the README's '24,278' is a digit swap). `text` = `full_text`; `source_record_id` = `essay_id` (unique across both files); `prompt` = `assignment`, the assignment wording the student answered (empty for the 28 essays whose assignment is 'NA' in the release); `meta` = every other column as released (score, set, pubpriv, assignment, prompt_name, economically_disadvantaged, student_disability_status, ell_status, race_ethnicity, gender, grade_level, essay_word_count, task; 'NA' strings kept), plus `source_text_file` = the PDF of the prompt's source texts in ASAP_2_source_texts.zip, and `persuade_2_essay_id_comp` = the list of essay_id_comp of the PERSUADE 2.0 essays with the same text (whitespace-collapsed exact match against both official PERSUADE 2.0 files; 12,873 essays — the README says 12,871 come from PERSUADE 2.0 —, which are also in this corpus as subset persuade_2 under CC BY-NC-SA 4.0). Left out: the source texts (ASAP_2_source_texts.zip, 7 PDFs of the test screens: reading passages, assignment and editing checklist). Five are page screenshots without a text layer, and the passages are third-party articles reprinted in the tests by permission with their own notices ('copyright (c) 2013 by Highlights for Children, Inc. ... Used by permission'; 'Copyright (c) 2004 by Mother Jones ... All Rights Reserved'; Cricket Media: 'Any commercial use or distribution of material without permission is strictly prohibited'), which the repo's CC BY 4.0 statement cannot cover; the file name is kept per essay. The scoring rubric (docx) is not an essay. Dates: the release gives none. For the 12,873 essays also in PERSUADE 2.0 `date` = 2021, the same upper bound as the persuade_2 subset (Kaggle Feedback Prize 2021); the others have no date (essays from state writing tests; the essay-id prefixes AAAOPP13416, AAATRP14318 and AAAVUP14319 and the years some essays mention suggest 2016, 2018 and 2019 test administrations, an inference, not a published date). No author ids or URLs are released.

*Licence:* CC BY 4.0 (official README at the pinned commit: 'The data is provided under a Attribution 4.0 International (CC BY 4.0) license'). The 12,873 essays that are also in PERSUADE 2.0 were released there by the same group under CC BY-NC-SA 4.0; ASAP 2.0 releases them under CC BY 4.0. The source-text PDFs carry third-party copyright notices and are not included.

### `caselaw_access_project`

Caselaw Access Project (Harvard Law School Library). Caselaw Access Project, Harvard Law School Library Innovation Lab (static.case.law bulk export, 2024).

Every case of every volume zip listed in VolumesMetadata.json (40,622 volumes, 6,920,598 cases), one row per case (json/<case>.json in the zip): 6,920,587 rows; the 11 cases whose casebody has neither head matter nor opinion text are dropped by the builder as empty. Nothing is sampled or filtered. `text` = the case body as plain text: the head matter (caption, docket number, court, date, counsel and, in volumes that are not redacted, the reporter's syllabus or headnotes) followed by each opinion (majority, concurrences, dissents, ... in the release's order), separated by blank lines; each part is the release's plain-text field (casebody.head_matter, casebody.opinions[].text) with its line breaks (one per paragraph). meta.parts gives each part's kind (head_matter / opinion), opinion type and author, and its [start, end) character offsets in `text` (offsets of the normalised text, valid for the published column). Cases whose casebody has no opinion (e.g. table dispositions: 'Petition for allowance of appeal denied') have only their head matter. Harvard-scanned volumes are uncorrected OCR (provenance.source 'Harvard'; analysis.ocr_confidence); Fastcase volumes are publisher text. `title` = name_abbreviation; `date` = decision_date; `source_record_id` = '<reporter_slug>/<volume_folder>/<file_name>'; `source_url` = the case JSON on static.case.law; `genre` 'U.S. court opinion (case reporter)'. `meta` holds the case JSON's other fields unchanged: id, name, name_abbreviation, decision_date, docket_number, first/last page and page order, file_name, citations, court, jurisdiction, analysis (word and character counts, OCR confidence, PageRank, simhash, sha256), provenance, last_updated, cites_to (the citation graph), and casebody's judges, parties, attorneys and corrections; plus `volume` (the volume's VolumesMetadata.json entry) and `reporter` (its ReportersMetadata.json id, full_name, short_name). Left out: the per-case HTML files (the same casebody with markup), the volume-level metadata JSON files (their content is in meta), and the scan PDFs and ALTO/image tars the site also offers.

*Licence:* CC0 1.0 for the Caselaw Data (case.law Terms of Use, effective 2024-03-13). Court opinions are public-domain government edicts; head matter written by reporter editors is released under the same CC0 terms.

### `casimir`

CASIMIR: A Corpus of Scientific Articles Enhanced with Multiple Author-Integrated Revisions. Jourdan, Boudin, Hernandez & Dufour (LREC-COLING 2024).

casimir: one row per article version of all three splits (mapping_<split>.jsonl: train 12,488 articles / 43,615 versions, dev 1,561 / 5,481, test 1,597 / 5,579; the split names are those of the HF dataset config: train, validation (= the *_dev files), test; the README table swaps the validation and test article counts). The release has no document-level text: article_pairs_<split>.jsonl holds, for each pair of consecutive versions, the sentences of both versions aligned one to one (a sentence without a counterpart is paired with an empty side), each with its VILA block type (Title, Section, Paragraph, Abstract, List, ...), section, paragraph and sentence numbers. A version's text is rebuilt from one side of one of its pairs (as in the earlier build): that side's sentences in sentence-pair-index order (rows where the version has no sentence id are skipped), a new block whenever the (type, section, paragraph) of the sentence changes or the sentence is a Title or Section, sentences joined by one space, blocks by a blank line. A version in two pairs (with the previous and the next version) is rebuilt from the side with fewer leaked git word-diff markers ('[-...-]', '{+...+}', a CASIMIR artefact), on a tie from the pair with the previous version; meta.rebuilt_from says which. The sentence text is as released: VILA PDF extraction with the authors' clean-up (bibliography removed; equations, figures and tables replaced by [Equation] / [Figure] / [Table]; page and line numbers and hard line breaks removed), so it keeps extraction noise ('(cid:N)' glyph codes, ligatures, letter-spaced small caps such as 'F AIRNESS'); the paragraph layout is our reconstruction. Every version occurs in at least one of the 36,733 pairs, so every version has a text (54,675 rows). Rows are in mapping order (article by article), versions oldest first. `source_record_id` = the OpenReview revision id; `parent_id` = the forum (article) id; `title` = the version's own title in its metadata record, else the forum title; `date` = the version's OpenReview tcdate (creation time of that revision, UTC). `meta`: `id_forum`, `version` (1 = oldest), `n_versions`, `split`, `rebuilt_from` (pair, side, number of pairs the version occurs in, diff markers in the chosen text), `version_metadata` (the version's metadata.jsonl record as released) and `forum_metadata` (the forum's record as released: title, authors, authorids, abstract, keywords, TL;DR, venue, venueid, _bibtex, pdf, dates, ...). Left out: article_pairs_small_test / mapping_small_test (a 468-article subset of test, same pairs). casimir_reviews: every note of reviews.jsonl (185,992 notes on 14,358 of the forums: official reviews, meta-reviews, decisions, author and public comments, withdrawal notices), one row each, split 'full' (the file has no split; meta.article_split gives the article's split). `text` = the note's content fields other than `title`, in release order; a note with one such field gets its value alone (e.g. the 'comment' of a comment), otherwise each field is a block '<field name>: <value>' (names as released, e.g. 'summary_of_the_paper', 'rating'; list values joined by ', '), blocks separated by a blank line; a note whose fields are all blank except the title (4 comments typed into the title field) gets its title as text, and one note that is blank everywhere has no text (dropped by the builder). `title` = content.title; `date` = tcdate; `parent_id` = forum id; `author_id` = the note's signature (an anonymous reviewer, area chair, authors or user group id); `meta` = the note's other fields as released (cdate, tcdate, tmdate, mdate, ddate, pdate, number, replyto, readers, writers, signatures, invitation(s), ...) plus `content_fields` (the field names in text order) and `article_split`. Everything in the release is written by people (article authors, reviewers, chairs); nothing is model-generated.

*Licence:* MIT (HF dataset card). The article texts remain their authors' work (public OpenReview PDFs), and the forum notes their writers' (reviewers, chairs, authors); cite CASIMIR and the OpenReview forum.

### `casimir_reviews`

CASIMIR: A Corpus of Scientific Articles Enhanced with Multiple Author-Integrated Revisions. Jourdan, Boudin, Hernandez & Dufour (LREC-COLING 2024).

casimir: one row per article version of all three splits (mapping_<split>.jsonl: train 12,488 articles / 43,615 versions, dev 1,561 / 5,481, test 1,597 / 5,579; the split names are those of the HF dataset config: train, validation (= the *_dev files), test; the README table swaps the validation and test article counts). The release has no document-level text: article_pairs_<split>.jsonl holds, for each pair of consecutive versions, the sentences of both versions aligned one to one (a sentence without a counterpart is paired with an empty side), each with its VILA block type (Title, Section, Paragraph, Abstract, List, ...), section, paragraph and sentence numbers. A version's text is rebuilt from one side of one of its pairs (as in the earlier build): that side's sentences in sentence-pair-index order (rows where the version has no sentence id are skipped), a new block whenever the (type, section, paragraph) of the sentence changes or the sentence is a Title or Section, sentences joined by one space, blocks by a blank line. A version in two pairs (with the previous and the next version) is rebuilt from the side with fewer leaked git word-diff markers ('[-...-]', '{+...+}', a CASIMIR artefact), on a tie from the pair with the previous version; meta.rebuilt_from says which. The sentence text is as released: VILA PDF extraction with the authors' clean-up (bibliography removed; equations, figures and tables replaced by [Equation] / [Figure] / [Table]; page and line numbers and hard line breaks removed), so it keeps extraction noise ('(cid:N)' glyph codes, ligatures, letter-spaced small caps such as 'F AIRNESS'); the paragraph layout is our reconstruction. Every version occurs in at least one of the 36,733 pairs, so every version has a text (54,675 rows). Rows are in mapping order (article by article), versions oldest first. `source_record_id` = the OpenReview revision id; `parent_id` = the forum (article) id; `title` = the version's own title in its metadata record, else the forum title; `date` = the version's OpenReview tcdate (creation time of that revision, UTC). `meta`: `id_forum`, `version` (1 = oldest), `n_versions`, `split`, `rebuilt_from` (pair, side, number of pairs the version occurs in, diff markers in the chosen text), `version_metadata` (the version's metadata.jsonl record as released) and `forum_metadata` (the forum's record as released: title, authors, authorids, abstract, keywords, TL;DR, venue, venueid, _bibtex, pdf, dates, ...). Left out: article_pairs_small_test / mapping_small_test (a 468-article subset of test, same pairs). casimir_reviews: every note of reviews.jsonl (185,992 notes on 14,358 of the forums: official reviews, meta-reviews, decisions, author and public comments, withdrawal notices), one row each, split 'full' (the file has no split; meta.article_split gives the article's split). `text` = the note's content fields other than `title`, in release order; a note with one such field gets its value alone (e.g. the 'comment' of a comment), otherwise each field is a block '<field name>: <value>' (names as released, e.g. 'summary_of_the_paper', 'rating'; list values joined by ', '), blocks separated by a blank line; a note whose fields are all blank except the title (4 comments typed into the title field) gets its title as text, and one note that is blank everywhere has no text (dropped by the builder). `title` = content.title; `date` = tcdate; `parent_id` = forum id; `author_id` = the note's signature (an anonymous reviewer, area chair, authors or user group id); `meta` = the note's other fields as released (cdate, tcdate, tmdate, mdate, ddate, pdate, number, replyto, readers, writers, signatures, invitation(s), ...) plus `content_fields` (the field names in text order) and `article_split`. Everything in the release is written by people (article authors, reviewers, chairs); nothing is model-generated.

*Licence:* MIT (HF dataset card). The article texts remain their authors' work (public OpenReview PDFs), and the forum notes their writers' (reviewers, chairs, authors); cite CASIMIR and the OpenReview forum.

### `chatbot_arena_140k`

Chatbot Arena human preference votes (arena-human-preference-140k and -55k). Chiang, Zheng, Sheng, Angelopoulos, Li, Li, Zhang, Zhu, Jordan, Gonzalez & Stoica (ICML 2024).

Two releases, one subset each, each with its one split `train`; every vote is kept (all languages). Each row is one user's vote between two anonymous models answering the same conversation. Model outputs are governed by the providers' terms, so `messages` holds the user turns with each assistant turn's `content` set to null, and the rows carry the licence of the user prompts. `model` is null because two models answered; their names and the vote are in `meta`. `chatbot_arena_140k` (lmarena-ai/arena-human-preference-140k, 135,634 votes, 2025-04 to 2025-07, CC BY 4.0): `messages` is the conversation of the vote (`conversation_a`; its user turns equal `conversation_b`'s, checked on every row; all content items are text, joined if there are several). Votes after the first in a session (`evaluation_order` > 1) were cast on a conversation that continued earlier ones; when the release's `full_conversation` (the whole session) has other user turns than `messages`, they are kept in `meta.full_conversation_user_turns` (27,319 rows). `meta` keeps `id`, `model_a`, `model_b`, `winner`, `evaluation_session_id`, `evaluation_order`, `conv_metadata`, `category_tag`, `language`, `is_code`, `timestamp` as released, and `conversation_b_num_tokens` (the release's per-message token counts of conversation B). `date` is `timestamp`. `chatbot_arena_55k` (lmarena-ai/arena-human-preference-55k, 57,477 battles of 2023 to early 2024; the training set of the Kaggle 'LMSYS - Chatbot Arena Human Preference Predictions' competition, whose test set was never released; Apache-2.0): `prompt` in the CSV is a JSON list of the user turns, `response_a` / `response_b` the answers (left out); `meta` keeps `id`, `model_a`, `model_b`, `winner_model_a`, `winner_model_b`, `winner_tie`. The 55k release has no dates, so `date` is null.

*Licence:* arena-human-preference-140k: "User prompts are licensed under CC-BY-4.0, and model outputs are governed by the terms of use set by the respective model providers" (README; card tag cc-by-4.0). arena-human-preference-55k: card tag Apache-2.0, no further terms. Model outputs are not included. May contain unsafe or offensive text.

### `chatbot_arena_55k`

Chatbot Arena human preference votes (arena-human-preference-140k and -55k). Chiang, Zheng, Sheng, Angelopoulos, Li, Li, Zhang, Zhu, Jordan, Gonzalez & Stoica (ICML 2024).

Two releases, one subset each, each with its one split `train`; every vote is kept (all languages). Each row is one user's vote between two anonymous models answering the same conversation. Model outputs are governed by the providers' terms, so `messages` holds the user turns with each assistant turn's `content` set to null, and the rows carry the licence of the user prompts. `model` is null because two models answered; their names and the vote are in `meta`. `chatbot_arena_140k` (lmarena-ai/arena-human-preference-140k, 135,634 votes, 2025-04 to 2025-07, CC BY 4.0): `messages` is the conversation of the vote (`conversation_a`; its user turns equal `conversation_b`'s, checked on every row; all content items are text, joined if there are several). Votes after the first in a session (`evaluation_order` > 1) were cast on a conversation that continued earlier ones; when the release's `full_conversation` (the whole session) has other user turns than `messages`, they are kept in `meta.full_conversation_user_turns` (27,319 rows). `meta` keeps `id`, `model_a`, `model_b`, `winner`, `evaluation_session_id`, `evaluation_order`, `conv_metadata`, `category_tag`, `language`, `is_code`, `timestamp` as released, and `conversation_b_num_tokens` (the release's per-message token counts of conversation B). `date` is `timestamp`. `chatbot_arena_55k` (lmarena-ai/arena-human-preference-55k, 57,477 battles of 2023 to early 2024; the training set of the Kaggle 'LMSYS - Chatbot Arena Human Preference Predictions' competition, whose test set was never released; Apache-2.0): `prompt` in the CSV is a JSON list of the user turns, `response_a` / `response_b` the answers (left out); `meta` keeps `id`, `model_a`, `model_b`, `winner_model_a`, `winner_model_b`, `winner_tie`. The 55k release has no dates, so `date` is null.

*Licence:* arena-human-preference-140k: "User prompts are licensed under CC-BY-4.0, and model outputs are governed by the terms of use set by the respective model providers" (README; card tag cc-by-4.0). arena-human-preference-55k: card tag Apache-2.0, no further terms. Model outputs are not included. May contain unsafe or offensive text.

### `chatgpt_detector_bias`

GPT Detectors Are Biased Against Non-Native English Writers (human-written data). Liang, Yuksekgonul, Mao, Wu & Zou (Patterns 2023).

The release has 10 subsets (925 texts), each a folder with data.json, name.json and per-detector score files. Included: every text of the 4 human-written subsets (name.json kind 'Human-Written'), 394 texts: Human_Data/TOEFL_real_91 (91 TOEFL essays by Chinese learners from the forum toefl.zhan.com, year <= 2020 per the paper), CollegeEssay_real_70 (70 US college admission essays from the prepscholar.com and collegeessayguy.com blogs), CS224N_real_145 (145 Stanford CS224N Winter 2021 project titles and abstracts, text 'Title: ...\nAbstract: ...' as released) and HewlettStudentEssay_real_88 (88 US 8th-grade essays from the Hewlett Foundation ASAP data). Left out, because they are AI-generated or AI-edited (name.json kind 'AI-Generated'), 531 texts: TOEFL_gpt4polished_91 (91, the TOEFL essays polished by GPT-4) and HewlettStudentEssay_GPTsimplify_88 (88, Hewlett essays simplified by GPT-4), both filed under Human_Data in the release, and GPT_Data/CS224N_gpt3_145 (145), CS224N_gpt3PromptEng_145 (145), CollegeEssay_gpt3_31 (31), CollegeEssay_gpt3PromptEng_31 (31). Code/ is not data. Nothing else is filtered (exact duplicates inside TOEFL_real_91 are kept as released). One row per data.json entry, in file order; `source_record_id` = '<subset>:<0-based index>'; `text` = 'document'; no titles, prompts, authors or URLs are released (`prompt` is empty). `genre` per subset. `date` is an upper bound, not the year of writing: TOEFL 2020 (paper: 'year<=2020'), CS224N 2021 (Winter 2021 quarter), Hewlett 2012 (the ASAP data were published in 2012); none for the college essays (collection dates undocumented). `meta`: subset, folder, name.json, 'type' (the college named for a CollegeEssay text), and detector_scores = for each detector file of the subset (Crossplag, GPTZero, HFOpenAI, OriginalityAI, Quil, Sapling, ZeroGPT; CS224N_real_145 has no Quil or Sapling file) the scalar fields of its entry for the text ('score', and GPTZero's perplexity, burstiness and probabilities; detector files list the texts in data.json order). In 6 CS224N_real_145 entries of GPTZero.json the scored text is the placeholder 'Title: ..\nAbstract: Project summaries unavailable', not the released abstract; those scores carry scored_document_differs = true. The detectors' raw API responses ('result_full': per-sentence HTML and the text again) are not kept.

*Licence:* MIT (README badge of the release and the authors' HF card; the GitHub repo has no LICENSE file; Zenodo lists 'other-open'). The texts were collected by the authors from third parties whose terms the release does not state: TOEFL essays from the forum toefl.zhan.com, college essays from the prepscholar.com and collegeessayguy.com blogs, abstracts from the Stanford CS224N project page, 8th-grade essays from the Hewlett Foundation ASAP competition data.

### `civil_comments`

Civil Comments (Jigsaw Unintended Bias in Toxicity Classification). Borkan, Dixon, Sorensen, Thain & Vasserman, Nuanced Metrics for Measuring Unintended Bias with Real Data for Text Classification (WWW 2019 companion).

Every comment of every split of the HF release: train 1,804,874, validation 97,320 (Kaggle test_public_expanded), test 97,320 (Kaggle test_private_expanded); 1,999,514 in all. One row per parquet row; `text` = `text` as released. `meta` = the seven Jigsaw labels as released (fraction of crowd raters who marked the comment: toxicity, severe_toxicity, obscene, threat, insult, identity_attack, sexual_explicit). The HF release has no comment ids, article ids, dates or author ids: `source_record_id` = '<split>/<row number in the split>' (row order of the parquet files at the pinned revision, from 0). `date` = '2017', an upper bound (the comments were written 2015-2017 and the platform shut down in 2017), not the year of writing. `genre` = 'news-site reader comment'. Rows with empty text are counted by the builder. Not used: the platform's own figshare archive (CC0, train portion only, with article ids and timestamps), which is a different release. Nothing is left out of the HF release.

*Licence:* CC0 1.0 (HF card 'license: cc0-1.0'; Jigsaw: 'This dataset is released under CC0, as is the underlying comment text').

### `cmu_movie_summaries`

CMU Movie Summary Corpus. Bamman, O'Connor & Smith (ACL 2013).

Every plot summary of the release (plot_summaries.txt: 42,306 lines, 'Wikipedia movie id<TAB>summary'), one row each, in file order; no split ('full'). `text` = the summary as released (one paragraph; wiki-markup residue from the extraction kept). `source_record_id` = the Wikipedia movie id (page id; `source_url` = https://en.wikipedia.org/?curid=<id>); `title` = Freebase movie name; `date` = 2012-11-02, the date of the Wikipedia dump the summaries were taken from (an upper bound: the summaries were written by Wikipedia editors at unknown earlier dates). `meta` = the movie's row of movie.metadata.tsv (Freebase id, name, release date, box office revenue, runtime, and the languages, countries and genres as {Freebase id: name} objects, parsed from the release's JSON strings; empty strings as released; meta.movie_metadata = false for the 99 summaries with no metadata row) and `characters`: the movie's rows of character.metadata.tsv in file order (character name, actor name, gender, date of birth, height, ethnicity Freebase id, age at release and the Freebase ids), each with the README's 'test data' persona labels of that character/actor map id where present (`tvtropes` from tvtropes.clusters.txt, `name_cluster` from name.clusters.txt). Left out: the metadata of the 39,534 movies without a summary (no text), and the optional CoreNLP-annotated copy of the summaries (a separate 628 MB download with the same texts).

*Licence:* CC BY-SA 3.0 US (README: 'All data is released under a Creative Commons Attribution-ShareAlike License'; the personas page links CC BY-SA 3.0 US). The summaries are English Wikipedia text (CC BY-SA). Attribution and share-alike.

### `co_storm_wildseek`

WildSeek (Co-STORM: Into the Unknown Unknowns). Jiang, Shao, Ma, Semnani & Lam (EMNLP 2024).

All 100 rows of data.json (no split in the release, so the split is `full`). Each row is a topic and the user's goal (`intent`) as typed into the STORM web application (storm.genie.stanford.edu) in 2024; the authors sampled them from 6,608 unique topic-goal pairs (paper App. A). `prompt` = 'Topic: <topic>' and 'Intent: <intent>' on two lines (the release has the two fields only; the labels are the release's field names). `category` = `domain` (the authors' topic taxonomy path); `meta` keeps topic, intent and domain unchanged. The release has no ids; `source_record_id` = the row's index in data.json.

*Licence:* CC BY-SA 4.0 (dataset card, README and LICENSE).

### `common_pile_foodista`

Common Pile v0.1 (raw): Ubuntu IRC, Foodista, Regulations.gov, USGPO, GitHub Archive. Kandpal, Lester, Raffel, Majstorovic, Biderman, Abbasi, Soldaini, Shippole, Cooper, Skowron, Longpre, Sutawika, Albalak, Xu, Penedo, et al. (arXiv 2025).

Every document of every document shard of five raw Common Pile v0.1 repos, one subset each: common_pile_ubuntu_irc (v0/documents, 8 shards), common_pile_foodista (v0/documents, 5), common_pile_regulations (v0, 7), common_pile_usgpo (data, 78), common_pile_github_archive (gharchive/v0/documents, 2,000). Split `train` (each repo's only split). These are the 'raw' repos, not the *_filtered versions used to train Comma v0.1 (those drop documents and redact PII). Not used: the repos' raw/ folders (copies before text extraction: IRC log files, Foodista HTML, GitHub raw threads), github_archive's licenses.jsonl (the repository licence table) and usgpo's licensed_pile_log.txt. One row per document, in shard order; `source_record_id` = '<repo>/<document id>' (e.g. 'ubuntu_irc/2015-04-17-#ubuntu-ir'; the repo name keeps ids unique across the five subsets); `text` = `text` as released (IRC: '[HH:MM] <nick> message' lines; Foodista: title, 'By:' and 'Published:' header lines, page text and comments; GitHub: the thread's posts converted from Markdown by Common Pile); `title` = `title` (regulations, usgpo; none for the others); `date` = `created` (ubuntu_irc: the log day; foodista: the page's published date; github_archive: the thread's creation time), `posted_date` (regulations) or `date` (usgpo); `source_url` = metadata.url; `genre` per subset ('IRC chat log (one channel, one day)', 'recipe and food web page', 'U.S. federal regulatory document', 'U.S. government publication', 'GitHub issue or pull request thread'). `meta` holds every other field as released: source, added, created / date / posted_date, document_type and agency (regulations), collection and author (usgpo), and the metadata dict (license, license_type and license_source for GitHub, url, authors, channel (IRC), repo (GitHub)), plus shard (the shard file). Documents with empty text (e.g. about a tenth of the regulations documents) are counted by the builder as dropped_empty. One id occurs twice in the release (regulations EPA-HQ-OAR-2002-0051-1991, two records of the same posted document that differ in `added` and url): the second occurrence in release order has source_record_id '...#2' and meta.id_occurrence 2 (ids are checked over all shards; raw/common_pile/duplicate_ids.json). Licence per document from metadata.license: 'Public Domain' -> public-domain (ubuntu_irc, regulations, usgpo), 'Creative Commons - Attribution - .../by/3.0/' -> cc-by-3.0 (foodista); GitHub's SPDX ids (either case) -> mit, apache-2.0, bsd-3-clause, cc0-1.0, cc-by-4.0, cc-by-sa-4.0, ... where the corpus has that licence id. GitHub documents under a licence the corpus has no id for (BSD-2-Clause, ISC, Unlicense, BSL-1.0, Zlib, WTFPL, MIT-0, 0BSD, Artistic-2.0, ...) are left out and counted per licence (raw/common_pile/_unmapped_licences/), because a row may only carry its own licence: 971,501 of the 30,318,774 GitHub threads (BSD-2-Clause 369,658, Unlicense 189,027, ISC 136,884, BSL-1.0 47,551, Zlib 42,019, WTFPL 41,548, MIT-0 25,960, 0BSD 19,532, Artistic-2.0 17,664, Vim 15,054, UPL-1.0 12,429, ECL-2.0 11,718, BSD-3-Clause-Clear 10,025, MS-PL 9,586, AFL-3.0 5,801, LPPL-1.3c 5,615, PostgreSQL 4,064, NCSA 3,200, BSD-4-Clause 2,629, MulanPSL-2.0 1,537; SPDX ids counted case-insensitively). Nothing else is filtered. Rows: ubuntu_irc 329,113 (+2 empty), foodista 72,034 (+56 empty), regulations 222,980 (+2,216 empty), usgpo 2,732,677, github_archive 29,347,255 (+18 empty), matching the documents counts in the repos' READMEs. Common Pile's own processing applies: it removed known bots' messages (IRC) and bot comments (GitHub), but automated posts remain (e.g. uptime-monitor issues); texts after 2022-11-30 (pre_chatgpt false) may contain machine-written text.

*Licence:* Per document, from Common Pile's metadata.license: Ubuntu IRC logs, Regulations.gov and USGPO documents are public domain (U.S. government works; IRC logs released into the public domain); Foodista content is CC BY 3.0; GitHub threads carry their repository's licence (GitHub terms: issue and PR text inherits the repository licence), as recorded by Common Pile. Common Pile warns that licence metadata can be wrong (licence laundering).

### `common_pile_github_archive`

Common Pile v0.1 (raw): Ubuntu IRC, Foodista, Regulations.gov, USGPO, GitHub Archive. Kandpal, Lester, Raffel, Majstorovic, Biderman, Abbasi, Soldaini, Shippole, Cooper, Skowron, Longpre, Sutawika, Albalak, Xu, Penedo, et al. (arXiv 2025).

Every document of every document shard of five raw Common Pile v0.1 repos, one subset each: common_pile_ubuntu_irc (v0/documents, 8 shards), common_pile_foodista (v0/documents, 5), common_pile_regulations (v0, 7), common_pile_usgpo (data, 78), common_pile_github_archive (gharchive/v0/documents, 2,000). Split `train` (each repo's only split). These are the 'raw' repos, not the *_filtered versions used to train Comma v0.1 (those drop documents and redact PII). Not used: the repos' raw/ folders (copies before text extraction: IRC log files, Foodista HTML, GitHub raw threads), github_archive's licenses.jsonl (the repository licence table) and usgpo's licensed_pile_log.txt. One row per document, in shard order; `source_record_id` = '<repo>/<document id>' (e.g. 'ubuntu_irc/2015-04-17-#ubuntu-ir'; the repo name keeps ids unique across the five subsets); `text` = `text` as released (IRC: '[HH:MM] <nick> message' lines; Foodista: title, 'By:' and 'Published:' header lines, page text and comments; GitHub: the thread's posts converted from Markdown by Common Pile); `title` = `title` (regulations, usgpo; none for the others); `date` = `created` (ubuntu_irc: the log day; foodista: the page's published date; github_archive: the thread's creation time), `posted_date` (regulations) or `date` (usgpo); `source_url` = metadata.url; `genre` per subset ('IRC chat log (one channel, one day)', 'recipe and food web page', 'U.S. federal regulatory document', 'U.S. government publication', 'GitHub issue or pull request thread'). `meta` holds every other field as released: source, added, created / date / posted_date, document_type and agency (regulations), collection and author (usgpo), and the metadata dict (license, license_type and license_source for GitHub, url, authors, channel (IRC), repo (GitHub)), plus shard (the shard file). Documents with empty text (e.g. about a tenth of the regulations documents) are counted by the builder as dropped_empty. One id occurs twice in the release (regulations EPA-HQ-OAR-2002-0051-1991, two records of the same posted document that differ in `added` and url): the second occurrence in release order has source_record_id '...#2' and meta.id_occurrence 2 (ids are checked over all shards; raw/common_pile/duplicate_ids.json). Licence per document from metadata.license: 'Public Domain' -> public-domain (ubuntu_irc, regulations, usgpo), 'Creative Commons - Attribution - .../by/3.0/' -> cc-by-3.0 (foodista); GitHub's SPDX ids (either case) -> mit, apache-2.0, bsd-3-clause, cc0-1.0, cc-by-4.0, cc-by-sa-4.0, ... where the corpus has that licence id. GitHub documents under a licence the corpus has no id for (BSD-2-Clause, ISC, Unlicense, BSL-1.0, Zlib, WTFPL, MIT-0, 0BSD, Artistic-2.0, ...) are left out and counted per licence (raw/common_pile/_unmapped_licences/), because a row may only carry its own licence: 971,501 of the 30,318,774 GitHub threads (BSD-2-Clause 369,658, Unlicense 189,027, ISC 136,884, BSL-1.0 47,551, Zlib 42,019, WTFPL 41,548, MIT-0 25,960, 0BSD 19,532, Artistic-2.0 17,664, Vim 15,054, UPL-1.0 12,429, ECL-2.0 11,718, BSD-3-Clause-Clear 10,025, MS-PL 9,586, AFL-3.0 5,801, LPPL-1.3c 5,615, PostgreSQL 4,064, NCSA 3,200, BSD-4-Clause 2,629, MulanPSL-2.0 1,537; SPDX ids counted case-insensitively). Nothing else is filtered. Rows: ubuntu_irc 329,113 (+2 empty), foodista 72,034 (+56 empty), regulations 222,980 (+2,216 empty), usgpo 2,732,677, github_archive 29,347,255 (+18 empty), matching the documents counts in the repos' READMEs. Common Pile's own processing applies: it removed known bots' messages (IRC) and bot comments (GitHub), but automated posts remain (e.g. uptime-monitor issues); texts after 2022-11-30 (pre_chatgpt false) may contain machine-written text.

*Licence:* Per document, from Common Pile's metadata.license: Ubuntu IRC logs, Regulations.gov and USGPO documents are public domain (U.S. government works; IRC logs released into the public domain); Foodista content is CC BY 3.0; GitHub threads carry their repository's licence (GitHub terms: issue and PR text inherits the repository licence), as recorded by Common Pile. Common Pile warns that licence metadata can be wrong (licence laundering).

### `common_pile_regulations`

Common Pile v0.1 (raw): Ubuntu IRC, Foodista, Regulations.gov, USGPO, GitHub Archive. Kandpal, Lester, Raffel, Majstorovic, Biderman, Abbasi, Soldaini, Shippole, Cooper, Skowron, Longpre, Sutawika, Albalak, Xu, Penedo, et al. (arXiv 2025).

Every document of every document shard of five raw Common Pile v0.1 repos, one subset each: common_pile_ubuntu_irc (v0/documents, 8 shards), common_pile_foodista (v0/documents, 5), common_pile_regulations (v0, 7), common_pile_usgpo (data, 78), common_pile_github_archive (gharchive/v0/documents, 2,000). Split `train` (each repo's only split). These are the 'raw' repos, not the *_filtered versions used to train Comma v0.1 (those drop documents and redact PII). Not used: the repos' raw/ folders (copies before text extraction: IRC log files, Foodista HTML, GitHub raw threads), github_archive's licenses.jsonl (the repository licence table) and usgpo's licensed_pile_log.txt. One row per document, in shard order; `source_record_id` = '<repo>/<document id>' (e.g. 'ubuntu_irc/2015-04-17-#ubuntu-ir'; the repo name keeps ids unique across the five subsets); `text` = `text` as released (IRC: '[HH:MM] <nick> message' lines; Foodista: title, 'By:' and 'Published:' header lines, page text and comments; GitHub: the thread's posts converted from Markdown by Common Pile); `title` = `title` (regulations, usgpo; none for the others); `date` = `created` (ubuntu_irc: the log day; foodista: the page's published date; github_archive: the thread's creation time), `posted_date` (regulations) or `date` (usgpo); `source_url` = metadata.url; `genre` per subset ('IRC chat log (one channel, one day)', 'recipe and food web page', 'U.S. federal regulatory document', 'U.S. government publication', 'GitHub issue or pull request thread'). `meta` holds every other field as released: source, added, created / date / posted_date, document_type and agency (regulations), collection and author (usgpo), and the metadata dict (license, license_type and license_source for GitHub, url, authors, channel (IRC), repo (GitHub)), plus shard (the shard file). Documents with empty text (e.g. about a tenth of the regulations documents) are counted by the builder as dropped_empty. One id occurs twice in the release (regulations EPA-HQ-OAR-2002-0051-1991, two records of the same posted document that differ in `added` and url): the second occurrence in release order has source_record_id '...#2' and meta.id_occurrence 2 (ids are checked over all shards; raw/common_pile/duplicate_ids.json). Licence per document from metadata.license: 'Public Domain' -> public-domain (ubuntu_irc, regulations, usgpo), 'Creative Commons - Attribution - .../by/3.0/' -> cc-by-3.0 (foodista); GitHub's SPDX ids (either case) -> mit, apache-2.0, bsd-3-clause, cc0-1.0, cc-by-4.0, cc-by-sa-4.0, ... where the corpus has that licence id. GitHub documents under a licence the corpus has no id for (BSD-2-Clause, ISC, Unlicense, BSL-1.0, Zlib, WTFPL, MIT-0, 0BSD, Artistic-2.0, ...) are left out and counted per licence (raw/common_pile/_unmapped_licences/), because a row may only carry its own licence: 971,501 of the 30,318,774 GitHub threads (BSD-2-Clause 369,658, Unlicense 189,027, ISC 136,884, BSL-1.0 47,551, Zlib 42,019, WTFPL 41,548, MIT-0 25,960, 0BSD 19,532, Artistic-2.0 17,664, Vim 15,054, UPL-1.0 12,429, ECL-2.0 11,718, BSD-3-Clause-Clear 10,025, MS-PL 9,586, AFL-3.0 5,801, LPPL-1.3c 5,615, PostgreSQL 4,064, NCSA 3,200, BSD-4-Clause 2,629, MulanPSL-2.0 1,537; SPDX ids counted case-insensitively). Nothing else is filtered. Rows: ubuntu_irc 329,113 (+2 empty), foodista 72,034 (+56 empty), regulations 222,980 (+2,216 empty), usgpo 2,732,677, github_archive 29,347,255 (+18 empty), matching the documents counts in the repos' READMEs. Common Pile's own processing applies: it removed known bots' messages (IRC) and bot comments (GitHub), but automated posts remain (e.g. uptime-monitor issues); texts after 2022-11-30 (pre_chatgpt false) may contain machine-written text.

*Licence:* Per document, from Common Pile's metadata.license: Ubuntu IRC logs, Regulations.gov and USGPO documents are public domain (U.S. government works; IRC logs released into the public domain); Foodista content is CC BY 3.0; GitHub threads carry their repository's licence (GitHub terms: issue and PR text inherits the repository licence), as recorded by Common Pile. Common Pile warns that licence metadata can be wrong (licence laundering).

### `common_pile_ubuntu_irc`

Common Pile v0.1 (raw): Ubuntu IRC, Foodista, Regulations.gov, USGPO, GitHub Archive. Kandpal, Lester, Raffel, Majstorovic, Biderman, Abbasi, Soldaini, Shippole, Cooper, Skowron, Longpre, Sutawika, Albalak, Xu, Penedo, et al. (arXiv 2025).

Every document of every document shard of five raw Common Pile v0.1 repos, one subset each: common_pile_ubuntu_irc (v0/documents, 8 shards), common_pile_foodista (v0/documents, 5), common_pile_regulations (v0, 7), common_pile_usgpo (data, 78), common_pile_github_archive (gharchive/v0/documents, 2,000). Split `train` (each repo's only split). These are the 'raw' repos, not the *_filtered versions used to train Comma v0.1 (those drop documents and redact PII). Not used: the repos' raw/ folders (copies before text extraction: IRC log files, Foodista HTML, GitHub raw threads), github_archive's licenses.jsonl (the repository licence table) and usgpo's licensed_pile_log.txt. One row per document, in shard order; `source_record_id` = '<repo>/<document id>' (e.g. 'ubuntu_irc/2015-04-17-#ubuntu-ir'; the repo name keeps ids unique across the five subsets); `text` = `text` as released (IRC: '[HH:MM] <nick> message' lines; Foodista: title, 'By:' and 'Published:' header lines, page text and comments; GitHub: the thread's posts converted from Markdown by Common Pile); `title` = `title` (regulations, usgpo; none for the others); `date` = `created` (ubuntu_irc: the log day; foodista: the page's published date; github_archive: the thread's creation time), `posted_date` (regulations) or `date` (usgpo); `source_url` = metadata.url; `genre` per subset ('IRC chat log (one channel, one day)', 'recipe and food web page', 'U.S. federal regulatory document', 'U.S. government publication', 'GitHub issue or pull request thread'). `meta` holds every other field as released: source, added, created / date / posted_date, document_type and agency (regulations), collection and author (usgpo), and the metadata dict (license, license_type and license_source for GitHub, url, authors, channel (IRC), repo (GitHub)), plus shard (the shard file). Documents with empty text (e.g. about a tenth of the regulations documents) are counted by the builder as dropped_empty. One id occurs twice in the release (regulations EPA-HQ-OAR-2002-0051-1991, two records of the same posted document that differ in `added` and url): the second occurrence in release order has source_record_id '...#2' and meta.id_occurrence 2 (ids are checked over all shards; raw/common_pile/duplicate_ids.json). Licence per document from metadata.license: 'Public Domain' -> public-domain (ubuntu_irc, regulations, usgpo), 'Creative Commons - Attribution - .../by/3.0/' -> cc-by-3.0 (foodista); GitHub's SPDX ids (either case) -> mit, apache-2.0, bsd-3-clause, cc0-1.0, cc-by-4.0, cc-by-sa-4.0, ... where the corpus has that licence id. GitHub documents under a licence the corpus has no id for (BSD-2-Clause, ISC, Unlicense, BSL-1.0, Zlib, WTFPL, MIT-0, 0BSD, Artistic-2.0, ...) are left out and counted per licence (raw/common_pile/_unmapped_licences/), because a row may only carry its own licence: 971,501 of the 30,318,774 GitHub threads (BSD-2-Clause 369,658, Unlicense 189,027, ISC 136,884, BSL-1.0 47,551, Zlib 42,019, WTFPL 41,548, MIT-0 25,960, 0BSD 19,532, Artistic-2.0 17,664, Vim 15,054, UPL-1.0 12,429, ECL-2.0 11,718, BSD-3-Clause-Clear 10,025, MS-PL 9,586, AFL-3.0 5,801, LPPL-1.3c 5,615, PostgreSQL 4,064, NCSA 3,200, BSD-4-Clause 2,629, MulanPSL-2.0 1,537; SPDX ids counted case-insensitively). Nothing else is filtered. Rows: ubuntu_irc 329,113 (+2 empty), foodista 72,034 (+56 empty), regulations 222,980 (+2,216 empty), usgpo 2,732,677, github_archive 29,347,255 (+18 empty), matching the documents counts in the repos' READMEs. Common Pile's own processing applies: it removed known bots' messages (IRC) and bot comments (GitHub), but automated posts remain (e.g. uptime-monitor issues); texts after 2022-11-30 (pre_chatgpt false) may contain machine-written text.

*Licence:* Per document, from Common Pile's metadata.license: Ubuntu IRC logs, Regulations.gov and USGPO documents are public domain (U.S. government works; IRC logs released into the public domain); Foodista content is CC BY 3.0; GitHub threads carry their repository's licence (GitHub terms: issue and PR text inherits the repository licence), as recorded by Common Pile. Common Pile warns that licence metadata can be wrong (licence laundering).

### `common_pile_usgpo`

Common Pile v0.1 (raw): Ubuntu IRC, Foodista, Regulations.gov, USGPO, GitHub Archive. Kandpal, Lester, Raffel, Majstorovic, Biderman, Abbasi, Soldaini, Shippole, Cooper, Skowron, Longpre, Sutawika, Albalak, Xu, Penedo, et al. (arXiv 2025).

Every document of every document shard of five raw Common Pile v0.1 repos, one subset each: common_pile_ubuntu_irc (v0/documents, 8 shards), common_pile_foodista (v0/documents, 5), common_pile_regulations (v0, 7), common_pile_usgpo (data, 78), common_pile_github_archive (gharchive/v0/documents, 2,000). Split `train` (each repo's only split). These are the 'raw' repos, not the *_filtered versions used to train Comma v0.1 (those drop documents and redact PII). Not used: the repos' raw/ folders (copies before text extraction: IRC log files, Foodista HTML, GitHub raw threads), github_archive's licenses.jsonl (the repository licence table) and usgpo's licensed_pile_log.txt. One row per document, in shard order; `source_record_id` = '<repo>/<document id>' (e.g. 'ubuntu_irc/2015-04-17-#ubuntu-ir'; the repo name keeps ids unique across the five subsets); `text` = `text` as released (IRC: '[HH:MM] <nick> message' lines; Foodista: title, 'By:' and 'Published:' header lines, page text and comments; GitHub: the thread's posts converted from Markdown by Common Pile); `title` = `title` (regulations, usgpo; none for the others); `date` = `created` (ubuntu_irc: the log day; foodista: the page's published date; github_archive: the thread's creation time), `posted_date` (regulations) or `date` (usgpo); `source_url` = metadata.url; `genre` per subset ('IRC chat log (one channel, one day)', 'recipe and food web page', 'U.S. federal regulatory document', 'U.S. government publication', 'GitHub issue or pull request thread'). `meta` holds every other field as released: source, added, created / date / posted_date, document_type and agency (regulations), collection and author (usgpo), and the metadata dict (license, license_type and license_source for GitHub, url, authors, channel (IRC), repo (GitHub)), plus shard (the shard file). Documents with empty text (e.g. about a tenth of the regulations documents) are counted by the builder as dropped_empty. One id occurs twice in the release (regulations EPA-HQ-OAR-2002-0051-1991, two records of the same posted document that differ in `added` and url): the second occurrence in release order has source_record_id '...#2' and meta.id_occurrence 2 (ids are checked over all shards; raw/common_pile/duplicate_ids.json). Licence per document from metadata.license: 'Public Domain' -> public-domain (ubuntu_irc, regulations, usgpo), 'Creative Commons - Attribution - .../by/3.0/' -> cc-by-3.0 (foodista); GitHub's SPDX ids (either case) -> mit, apache-2.0, bsd-3-clause, cc0-1.0, cc-by-4.0, cc-by-sa-4.0, ... where the corpus has that licence id. GitHub documents under a licence the corpus has no id for (BSD-2-Clause, ISC, Unlicense, BSL-1.0, Zlib, WTFPL, MIT-0, 0BSD, Artistic-2.0, ...) are left out and counted per licence (raw/common_pile/_unmapped_licences/), because a row may only carry its own licence: 971,501 of the 30,318,774 GitHub threads (BSD-2-Clause 369,658, Unlicense 189,027, ISC 136,884, BSL-1.0 47,551, Zlib 42,019, WTFPL 41,548, MIT-0 25,960, 0BSD 19,532, Artistic-2.0 17,664, Vim 15,054, UPL-1.0 12,429, ECL-2.0 11,718, BSD-3-Clause-Clear 10,025, MS-PL 9,586, AFL-3.0 5,801, LPPL-1.3c 5,615, PostgreSQL 4,064, NCSA 3,200, BSD-4-Clause 2,629, MulanPSL-2.0 1,537; SPDX ids counted case-insensitively). Nothing else is filtered. Rows: ubuntu_irc 329,113 (+2 empty), foodista 72,034 (+56 empty), regulations 222,980 (+2,216 empty), usgpo 2,732,677, github_archive 29,347,255 (+18 empty), matching the documents counts in the repos' READMEs. Common Pile's own processing applies: it removed known bots' messages (IRC) and bot comments (GitHub), but automated posts remain (e.g. uptime-monitor issues); texts after 2022-11-30 (pre_chatgpt false) may contain machine-written text.

*Licence:* Per document, from Common Pile's metadata.license: Ubuntu IRC logs, Regulations.gov and USGPO documents are public domain (U.S. government works; IRC logs released into the public domain); Foodista content is CC BY 3.0; GitHub threads carry their repository's licence (GitHub terms: issue and PR text inherits the repository licence), as recorded by Common Pile. Common Pile warns that licence metadata can be wrong (licence laundering).

### `congressional_hearings`

Congressional hearings (govinfo CHRG). U.S. Congress, House, Senate and Joint Committee hearings (U.S. Government Publishing Office, govinfo collection CHRG).

One row per hearing package of the CHRG collection listed in govinfo's sitemaps (one sitemap per hearing year, 1936-2026, the year of the first held date for all but a handful; a package is one printed hearing, or one part of a multi-part hearing: 'CHRG-<congress><h|s|j>hrg<jacket>'). Nothing is sampled or filtered. `text` = the package's HTML text rendition (GPO's text of the printed hearing: cover and title pages, committee roster, contents, opening statements, testimony, questions and answers, prepared statements, material submitted for the record, as printed) with the HTML wrapper removed: the content of its <pre> block, verbatim, with GPO's hard line breaks, indentation, page and section rules and bracketed editorial notes. The only changes: links are replaced by their text; e-mail addresses that govinfo's Cloudflare front end replaces by an '[email protected]' link are decoded back to the address printed in GPO's file; HTML character references (&amp; &lt; &gt; &#160; &middot; ...) are unescaped. GPO's typesetting codes in angle brackets are not HTML and are kept as printed: '<bullet>', '<GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT>', '<INF>'/'<SUP>' (sub/superscript), '<F-dash>', '<all>' (end of document), '<TEXT NOT AVAILABLE>', ... A package with several granules (the hearing plus addenda) gets the HTML files of all granules in MODS order, separated by a blank line. Some packages (mostly 1990s) have only the cover and title pages followed by a GPO placeholder ('<TEXT NOT AVAILABLE>', '[TEXT NOT AVAILABLE REFER TO PDF]', '<TEXT NOT AVAILABLE IN TIFF FORMAT>', '[NO TEXT AVAILABLE]', ...); they are kept, and every text containing such a placeholder is flagged meta.text_not_available = true (a few long hearings carry one for a single omitted insert). Result: 35,696 rows of the 45,612 packages (1,760 flagged text_not_available; 109 with several granules). The other 9,916 packages have no HTML rendition, so no text, and are counted by the builder as dropped_empty: 9,911 digitised historical hearings (MODS digitalOrigin 'reformatted digital'), which govinfo publishes as scanned PDF only (233 other digitised packages do have an HTML text and are kept); 4 born-digital packages whose MODS lists only a PDF rendition (CHRG-116hhrg43625, CHRG-117jhrg54192, CHRG-119jhrg61127, CHRG-119jhrg61803); and CHRG-105jhrg, a sitemap entry whose MODS and HTML both redirect to govinfo's error page. PDFs are not used. `title` = MODS titleInfo/title; `date` = the hearing's held date (MODS heldDate; the earliest when a hearing ran over several days — all are in meta.heldDate; dateIssued for the 4 packages whose MODS gives no heldDate, flagged meta.date_from = 'dateIssued'); `genre` from docClass: HHRG 'House committee hearing', SHRG 'Senate committee hearing', JHRG 'Joint committee hearing'; `source_record_id` = package id; `source_url` = the govinfo details page; `author_id` and `parent_id` empty. `meta` holds the package MODS record: every <extension> element under its MODS name (docClass, chamber, congress, session, number, heldDate, dateIngested, preferredCitation, congSerial, jacketId, searchTitle, congCommittee with authorityId and sub-committees, congMember with bioGuideId, party, state, role and names, witness, nominee, bill, law, USCode, agency, ...; elements that can repeat are lists), sudoc (SuDoc classification), dateIssued, extent (pages), digitalOrigin, sourceContentType, identifier (uri, local, former package identifier, ...), relatedItem (isReferencedBy: serial numbers, jacket citation), recordCreationDate, recordChangeDate, titleAlternative; granules (for the packages with constituent granules: each granule's accessId, title, partName, identifiers and extension, whose witness / congMember / congCommittee entries are also merged into the package-level lists); html_files (each HTML file with its HTTP status); sitemap_year and sitemap_lastmod.

*Licence:* U.S. Government work, not subject to copyright in the United States (17 U.S.C. 105); govinfo states no licence ('Public Domain & Copyright Notice': 'Government publications may contain copyrighted material which was used with permission of the copyright owner'). Hearing records print statements of non-federal witnesses and inserted third-party material (letters, articles, reports), which can carry their own copyright.

### `congressional_record`

Congressional Record (daily edition, govinfo CREC). U.S. Congress, Congressional Record, daily edition, vols. 140-172 (U.S. Government Publishing Office, govinfo collection CREC, 1994-2026).

One row per granule (govinfo's unit of the Record: a titled speech or debate section, a page of proceedings, an Extension of Remarks, a Daily Digest section, ...) of every daily issue govinfo lists (5,984 issues, 1994-01-25 to 2026-10-09), all four parts (House, Senate, Extensions of Remarks, Daily Digest). Nothing is sampled or filtered. `text` is the granule's HTML file (`html/<granule>.htm`) with the HTML wrapper removed: the content of its <pre> block, verbatim, with the hard line breaks of GPO's text rendition, including GPO's bracketed header lines ('[Congressional Record Volume ..]', '[Senate]', '[Page S123]', 'From the Congressional Record Online ..'), page markers ('[[Page S124]]'), '{time} 1415' stamps, speaker tags ('Mr. SMITH.'), editorial notes, inserted letters, bill text and tables. The only changes: <a href=..> links (GPO's auto-linked URLs) are replaced by their link text, and the few escaped characters (&gt; &lt; &amp; &quot; &#..;) are unescaped. GPO's typesetting codes in angle brackets are not HTML and are kept as printed in the file: '<bullet>' (marks Senate statements not delivered on the floor), '<GRAPHIC><TIFF>..', '<INF>', '<SUP>', '<gr-thn-eq>', '<plus-minus>', '<divide>', ... The builder's normalisation then collapses runs of spaces inside lines, so the release's indentation (two spaces start a paragraph; deeper indents mark block material) shows as one leading space and table columns lose their alignment. `title` = MODS titleInfo/title; `date` = MODS granuleDate (for Extensions of Remarks this is often the day the remarks were made, earlier than the issue date, which is in meta.package.dateIssued); `genre` from granuleClass: HOUSE 'House floor proceedings', SENATE 'Senate floor proceedings', EXTENSIONS 'Extensions of Remarks', DAILYDIGEST 'Daily Digest'; `parent_id` = package id (CREC-YYYY-MM-DD); `source_record_id` = granule id (MODS accessId = HTML file name); `source_url` = the govinfo HTML rendition; `author_id` = the member's Bioguide id for an Extension of Remarks with exactly one congMember, else empty (floor granules have many speakers). `meta` holds the granule's MODS record: the whole <extension> under its MODS names (granuleClass, subGranuleClass, chamber, pagePrefix, bookNumber, time from/to, congMember with bioGuideId, party, state, role and the parsed and authority names, congCommittee, bill, amendment, law, USCode, statuteAtLarge, congReport, congDoc, congVote, execCommunication, rin, associatedBills, ...), partName, identifier (preferred citation, congressional record citation, former granule identifier, ...), pages (start/end), extent, pdf_url, and package fields (id, dateIssued, volume, issue, congress, session, partNumber, dateIngested). Two parts are condensed: the roll-call list of a congVote (a few hundred congMember entries) is kept as congMember_byRole = {role: [bioGuideId, ...]} (names dropped; they follow from the Bioguide id), and the MODS <name> entries and otherFormat links of a granule are not repeated (they restate congMember / congCommittee and the URLs). 18 HTML files of 1994-1995 issues have no MODS granule entry (the quarterly Lobby List supplements, file names '..-Pgnull..'); they are included with title = their bracketed '[Lobby List: ..]' heading, genre 'Lobby List', date = issue date and meta.modsEntry = false. A MODS granule without an HTML file has no text and is counted by the builder as dropped_empty. PDFs are not used (the HTML holds the same text).

*Licence:* U.S. Government work, not subject to copyright in the United States (17 U.S.C. 105); govinfo states no licence. The Record reprints some third-party material (newspaper articles, letters, reports) inserted by members, which can carry its own copyright.

### `courtlistener`

CourtListener bulk opinions (Free Law Project), snapshot 2026-09-30. Free Law Project, CourtListener bulk data (snapshot 2026-09-30).

Every row of the opinions table of the 2026-09-30 snapshot, one corpus row per opinion (a case = an opinion cluster can hold several opinions: lead, concurrence, dissent, ...; `parent_id` = the cluster id): 10,903,928 opinions read, 10,901,752 rows; the 2,176 opinions with no text in any text field (or only empty markup) are dropped by the builder as empty. Text source: plain_text 2,711,136, xml_harvard 5,091,916, html_lawbox 1,490,019, html_columbia 939,451, html 622,516, html_anon_2020 46,597, html_with_citations 117. 7 opinions have no cluster and 11 no docket in the snapshot. Nothing is sampled or filtered: privacy-blocked cases (cluster/docket blocked = 't', hidden from search engines by CourtListener but published in the bulk file) and OCR-extracted texts are kept and flagged in meta. `text` = the opinion's best text field: plain_text when it has text (as released, e.g. the text layer of a court PDF, with its line breaks), else the first non-empty of html_columbia, html_lawbox, xml_harvard, html_anon_2020, html, xml_scan, converted to plain text (block elements p, div, blockquote, headings, list items, table rows, footnotes, <br> become line or paragraph breaks; other whitespace collapses as in HTML; <pre> keeps its layout; entities decoded; tags dropped, their text kept, including star-pagination marks such as '*1003' and footnote numbers), else html_with_citations (CourtListener's citator output, used only when no other field has text); meta.text_field says which field was used and meta.text_fields_chars gives the length of every non-empty text field. The text fields themselves are not repeated in meta. `date` = the cluster's date_filed; `title` = the cluster's case_name (else case_name_full, else the docket's case_name); `author_id` = the opinion's author_id (CourtListener person id) when set; `source_url` = the cluster's CourtListener page; `genre` 'U.S. court opinion'. `meta` = the opinion's other columns as released (id, date_created, date_modified, author_str, per_curiam, joined_by_str, type, sha1, page_count, download_url, local_path, extracted_by_ocr, author_id, cluster_id; empty values omitted) plus `cluster` (all non-empty columns of its opinion-clusters row: case names, judges, date_filed, precedential_status, source codes, citation_count, syllabus, headnotes, summary, headmatter, attorneys, ..., and `citations` = its rows of the citations table), `docket` (all non-empty columns of the cluster's docket) and `court` (id, full_name, short_name, citation_string, jurisdiction, url, start/end date, parent_court_id, in_use). An opinion whose cluster or docket is missing from the snapshot keeps null there. Overlap: CourtListener imported the Caselaw Access Project (cluster.source contains 'U'; xml_harvard set), Columbia and Lawbox collections, so many opinions here are also in the caselaw_access_project subset, usually with different formatting; both are kept as released (text_hash finds exact duplicates only). Rows are in file order (database order, not by id or date). CSV parsing: the reader joins pieces split at false row starts and counts unparseable starts (see out/_stats).

*Licence:* Free Law Project: the bulk data files are 'free of known copyright restrictions' (Public Domain Mark 1.0). Court opinions are public-domain government edicts; some imported fields (e.g. headnotes or summaries from commercial collections) are released by Free Law Project under the same terms.

### `cuad`

CUAD: Contract Understanding Atticus Dataset v1. Hendrycks, Burns, Chen & Ball (NeurIPS 2021 Datasets and Benchmarks).

Every contract of the release: all 510 (one row each). split = the paper's contract-level split as shipped in the GitHub data.zip: 408 contracts of train_separate_questions.json -> 'train', 102 of test.json -> 'test' (the two files partition the 510). `text` = the contract's full text from CUAD_v1/full_contract_txt/<title>.txt (the release's PDF-to-text conversion of the EDGAR exhibit; it equals the `context` of CUAD_v1.json for all 510 contracts). `source_record_id` = the CUAD title (the file stem: filer, filing date, form, exhibit number, contract type); `title` = the expert annotation 'Document Name-Answer' (the contract's name, e.g. 'DISTRIBUTOR AGREEMENT'); `date` = the filing date encoded in the file name (MM_DD_YYYY or YYYYMMDD); for the 27 contracts whose file name has no date it is '2021-03', the release date (an upper bound; meta.date_source says which); `genre` 'commercial contract (SEC EDGAR exhibit)'. `meta` holds the clause annotations as released: `qas` = the contract's 41 SQuAD-style questions from CUAD_v1.json (id, question, is_impossible, answers with text and answer_start; answer_start counts characters of the released text, before the builder's normalisation), `master_clauses` = the contract's row of master_clauses.csv (all 83 columns: per category the extracted clause text and the human answer, e.g. 'Agreement Date-Answer', 'Governing Law-Answer', Yes/No answers), and `pdf_file` (the matching PDF in full_contract_pdf/). The PDFs and the 28 label-group Excel files (the same labels by category) are not used. Nothing is left out.

*Licence:* CC BY 4.0 (Zenodo record licence; CUAD README: 'CUAD is licensed under the Creative Commons Attribution 4.0 (CC BY 4.0) license and free to the public for commercial and non-commercial use. The creators make no representations or warranties regarding the license status of the underlying contracts, which are publicly available and downloadable from EDGAR.')

### `djinni`

Djinni Recruitment Dataset (English job descriptions). Drushchak & Romanyshyn (UNLP 2024).

Every row of the release's only split (train, 141,897 job posts). `text` = `Long Description` (the job post body as released, with its Markdown-style *bold* markers and line breaks), `title` = `Position`, `source_record_id` = `id` (the release's UUID), `author_id` = `Company Name` (the posting company), `date` = the month of `Published` (YYYY-MM; the release gives only the first day of the publication month, Kyiv time). `meta` holds the other fields as released: `Company Name`, `Exp Years`, `Primary Keyword`, `English Level`, `Published`, `Long Description_lang` (the authors' language id, 'en' for every row) and `__index_level_0__` (the pandas index the release kept). Nothing is left out. The paper's other three repos (English candidate CVs, Ukrainian job descriptions, Ukrainian CVs) are separate releases and are not part of this source.

*Licence:* MIT (dataset card `license: mit`; paper Sec. 7). Keep the MIT notice.

### `ebg_stylometry`

Extended Brennan-Greenstadt and Brennan-Greenstadt adversarial stylometry corpora. Brennan, Afroz & Greenstadt (ACM TISSEC 2012); Brennan & Greenstadt (PETS 2009).

Two corpora of the lab's repository, one subset each, no official split ('full'). One row per text file, in path order. ebg_stylometry = every document of the Extended Brennan-Greenstadt corpus (jsan_resources/corpora/amt, 45 writers, 847 files): 699 pre-existing writing documents (<writer>_<nn>[_<k>...].txt), 45 obfuscation passages, 49 imitation passages (45 imitation_01 written to the McCarthy brief, 4 imitation_02 with an undocumented target) and 9 verification documents = 802 rows; the obfuscation and imitation passages were written by the same human writers as attacks on authorship recognition and are kept, labelled in meta.doc_type and meta.task. The 45 <writer>_demographics.txt files are not documents: their fields (age, language, gender, education, occupation, background) go to meta.demographics of every row of that writer, and the file as released to meta.demographics_raw. ebg_stylometry_bg = the original Brennan-Greenstadt corpus as the repository holds it (jsan_resources/corpora/drexel_1): the pre-existing writing of its 12 writers (102 documents). Left out: drexel_1/cm (11 excerpts of Cormac McCarthy's novels, the imitation target: copyrighted text by a third party, not the participants' writing and not covered by the repository licence); and the BG corpus's 12 obfuscation and 12 imitation passages, which are not in the repository (only in the lab's former website zip brennan-greenstadt-corpus.zip, which states no licence). `text` = the file decoded as UTF-8 (a leading byte-order mark removed); 11 BG files are not valid UTF-8: they were saved on a Mac, so their other non-ASCII bytes are decoded as Mac Roman (e.g. 'Böll', 'café'), meta.encoding = 'mac_roman+utf-8'. `author_id` = the writer's code (the directory name, e.g. 'aa'); `source_record_id` = 'amt/<writer>/<file>' or 'drexel_1/<writer>/<file>'. `date` = an upper bound, not the date of writing: 2011 for EBG (collected 2010-2011 on Mechanical Turk; the lab's archive of it is dated 2011-12-21; pre-existing documents can be older) and 2009 for BG (collected for the PETS 2009 paper). `meta` = corpus, writer, file, doc_type (pre_existing / obfuscation / imitation / verification, from the file name), task (a plain description of the writing task, from the TISSEC paper and the lab's participation brief), git blob SHA-1 and, for EBG, the demographics.

*Licence:* BSD 3-Clause (repository files '3-Clause BSD.txt' and license.txt: Copyright (c) 2013, Privacy Security and Automation Lab at Drexel University; README: released under AGPLv3 in 2011, BSD-3 since 2013). Keep the copyright notice and licence text. Writers agreed that their submissions would be added to a public research collection (the lab's participation brief).

### `ebg_stylometry_bg`

Extended Brennan-Greenstadt and Brennan-Greenstadt adversarial stylometry corpora. Brennan, Afroz & Greenstadt (ACM TISSEC 2012); Brennan & Greenstadt (PETS 2009).

Two corpora of the lab's repository, one subset each, no official split ('full'). One row per text file, in path order. ebg_stylometry = every document of the Extended Brennan-Greenstadt corpus (jsan_resources/corpora/amt, 45 writers, 847 files): 699 pre-existing writing documents (<writer>_<nn>[_<k>...].txt), 45 obfuscation passages, 49 imitation passages (45 imitation_01 written to the McCarthy brief, 4 imitation_02 with an undocumented target) and 9 verification documents = 802 rows; the obfuscation and imitation passages were written by the same human writers as attacks on authorship recognition and are kept, labelled in meta.doc_type and meta.task. The 45 <writer>_demographics.txt files are not documents: their fields (age, language, gender, education, occupation, background) go to meta.demographics of every row of that writer, and the file as released to meta.demographics_raw. ebg_stylometry_bg = the original Brennan-Greenstadt corpus as the repository holds it (jsan_resources/corpora/drexel_1): the pre-existing writing of its 12 writers (102 documents). Left out: drexel_1/cm (11 excerpts of Cormac McCarthy's novels, the imitation target: copyrighted text by a third party, not the participants' writing and not covered by the repository licence); and the BG corpus's 12 obfuscation and 12 imitation passages, which are not in the repository (only in the lab's former website zip brennan-greenstadt-corpus.zip, which states no licence). `text` = the file decoded as UTF-8 (a leading byte-order mark removed); 11 BG files are not valid UTF-8: they were saved on a Mac, so their other non-ASCII bytes are decoded as Mac Roman (e.g. 'Böll', 'café'), meta.encoding = 'mac_roman+utf-8'. `author_id` = the writer's code (the directory name, e.g. 'aa'); `source_record_id` = 'amt/<writer>/<file>' or 'drexel_1/<writer>/<file>'. `date` = an upper bound, not the date of writing: 2011 for EBG (collected 2010-2011 on Mechanical Turk; the lab's archive of it is dated 2011-12-21; pre-existing documents can be older) and 2009 for BG (collected for the PETS 2009 paper). `meta` = corpus, writer, file, doc_type (pre_existing / obfuscation / imitation / verification, from the file name), task (a plain description of the writing task, from the TISSEC paper and the lab's participation brief), git blob SHA-1 and, for EBG, the demographics.

*Licence:* BSD 3-Clause (repository files '3-Clause BSD.txt' and license.txt: Copyright (c) 2013, Privacy Security and Automation Lab at Drexel University; README: released under AGPLv3 in 2011, BSD-3 since 2013). Keep the copyright notice and licence text. Writers agreed that their submissions would be added to a public research collection (the lab's participation brief).

### `edgar_corpus`

EDGAR-CORPUS: Billions of Tokens Make The World Go Round. Loukas, Fergadiotis, Androutsopoulos & Malakasiotis (ECONLP 2021).

Every filing of the release: all 84 files (fiscal years 1993-2020 x the release's fixed train / validate / test partition; the card says the partition has no special meaning), 220,375 filings (train 176,289, validate 22,050, test 22,036), one row per filing; split = the release's own split name (train / validate / test). Nothing is sampled or filtered. `text` = the filing's item sections joined in the order of Form 10-K (release fields section_1, 1A, 1B, 2, 3, 4, 5, 6, 7, 7A, 8, 9, 9A, 9B, 10, 11, 12, 13, 14, 15), separated by a blank line. Each section as released opens with the filing's own item heading ('Item 1. Business', 'ITEM 7. MANAGEMENT'S DISCUSSION AND ANALYSIS ...'; captions and item numbering as the registrant printed them, which differ between years; checked: every non-empty section of the release begins with its 'Item'/'ITEM' heading), so no heading is added. Empty sections (the release stores an item the extractor did not find, or that the filing does not have, as an empty string) are left out; a filing whose 20 sections are all empty has no text and is counted by the builder as dropped_empty (1,607 filings: train 1,303, validate 158, test 146; 218,768 rows remain). The release has no fields for the cover page, Part headings, signatures or exhibits, so those are not in the text; the corpus authors stripped tables during extraction (edgar-crawler). Section texts are as released (the builder's normalisation only). `source_record_id` = the release's file name '<cik>_<fiscal year>.<htm|txt>' (unique across the release); `title` = None (the release has no company name; the filing is identified by cik and year in meta); `date` = the fiscal year ('year', the period the report covers, not the filing date: a 10-K is filed up to three months after the fiscal year ends); `author_id` = the CIK (the registrant); `source_url` = None (the release has no accession number, so no EDGAR link can be formed). `meta` = {cik, year, filename, hf_file and line (the release file and line), section_lengths: the character length of each of the 20 section fields as released, in item order (0 = empty)}. Not used: the authors' earlier Zenodo release (record 5528490, the same corpus in yearly zips without a partition).

*Licence:* Stated inconsistently: the HF card metadata and the loading script say apache-2.0 (card text: 'EDGAR data is publicly available'); the authors' Zenodo record of the same corpus (zenodo.org/records/5528490, 2021-09-26) says cc-by-4.0. Rows carry CC BY 4.0, the stricter reading (attribution to the corpus and its paper). The underlying 10-K filings are public SEC records written by the registrants.

### `eli5`

ELI5: Long Form Question Answering (KILT release). Fan, Jernite, Perez, Grangier, Weston & Auli (ACL 2019); Petroni et al., KILT (NAACL 2021).

Every question of the three official KILT ELI5 files: train (eli5-train-kilt.jsonl, 272,634), dev (eli5-dev-kilt.jsonl, 1,507; split `validation`) and test (eli5-test_without_answers-kilt.jsonl, 600; KILT withholds the test answers). The ELI5 authors host no processed data and point to KILT, so KILT is the official release. `prompt` (and the single user turn in `messages`) is the question title as KILT releases it (KILT removed the 'ELI5:' tag). `human_response` is the first answer of the release's `output` list (Reddit answers written by people; train answers carry their Reddit score); test rows have none. `meta.output` keeps the whole released `output` list unchanged: every answer (with `meta.score` in train) and, in dev, the Wikipedia provenance entries added by KILT annotators; `meta.kilt_meta` keeps the row-level KILT `meta` (partial Wikipedia evidence) where present. `source_record_id` is the Reddit thread id and `source_url` the thread. No date is given in the release. Five train questions are empty and are dropped by the builder.

*Licence:* KILT is MIT licensed (KILT README and LICENSE). Questions and answers are Reddit user content; Reddit's terms apply to the original posts.

### `elife_plos_lay_summaries`

PLOS and eLife lay summarisation corpora (Making Science Simple). Goldsack, Zhang, Lin & Scarton (EMNLP 2022).

Every article of both corpora and all three splits of the release: PLOS train 24,773 / validation 1,376 / test 1,376 and eLife train 4,346 / validation 241 / test 241 (files train.json, val.json, test.json in each zip; the HF loader calls val.json 'validation'), one row per article, PLOS before eLife, in file order. Nothing is filtered. The release stores each article as sentence lists, word-tokenised by the authors (spaces around punctuation, e.g. 'sequences .', '200 , 000', 'et al .'), without paragraph breaks; that text is kept as released. `text` = the article body: for each section, its heading (`headings`) as its own line, then the section's sentences joined by one space as one block; headings and blocks separated by a blank line. The abstract is not part of `text` (the HF loader prepends it to `article`); it is in `meta.abstract`. `meta.lay_summary` = the lay summary (`summary`: PLOS author summary written by the article's authors; eLife digest written by eLife's editors with the authors) and `meta.abstract` = the abstract, each as its sentences joined by one space, as the HF loader joins them; both human-written. `meta` also keeps `corpus` (plos / elife), `id`, `year`, `keywords` (list), `headings` (list) and, for the 1,691 PLOS articles that have it, `toc` (the journal's one-sentence table-of-contents blurb) as released. `title` = title; `date` = year; `source_record_id` = '<corpus>/<id>' (PLOS id = DOI suffix, eLife id = article number and version); `source_url` = https://doi.org/10.1371/<id> for PLOS and https://elifesciences.org/articles/<number>v<version> for eLife. No author ids are released.

*Licence:* CC BY 4.0 (the authors' HF dataset card: license cc-by-4.0). The articles are open-access PLOS and eLife papers published under CC BY 4.0; credit the article authors and the corpus.

### `europarl`

Europarl v7: European Parliament Proceedings Parallel Corpus 1996-2011 (source release). Koehn (MT Summit 2005).

Every speaker turn of every day file of the v7 source release (europarl.tgz: 187,072 files txt/<lang>/ep-YY-MM-DD[-NNN[-NN]].txt in 21 languages, sittings of 1996-04-15 to 2011-11-16; 2,831,433 SPEAKER turns), one row per turn per language. da, de, el, en, es, fr, it, nl, pt start in 1996-04, fi and sv in 1997-01, cs, et, hu, lt, lv, pl, sk, sl in 2006-01 and bg, ro in 2007-01 (later sittings come as one file per agenda item, so file counts are similar). Nothing is sampled or filtered; the 43,855 turns with no text are dropped by the builder, leaving 2,787,578 rows. `text` = the lines of the turn as released, a <P> (or </P>, <P ALIGN=..>, <BRK>) marker line written as a blank line, so paragraphs are separated by a blank line; the only change is that the release's few escaped characters (&amp; &lt; &gt; &quot; &#x..;) are unescaped. Procedural notes inside a turn ('(Applause)', '(The Minutes were approved)') and the language marker some versions put at the start of a translated speech ('(FI) Madam President, ..') are kept as released. Release errors are kept too: some voting-time turns hold text of several speakers, and 69,937 turns reuse a SPEAKER ID already used in the same chapter. `title` = the first heading line of the turn's chapter (the lines between a CHAPTER tag and its first SPEAKER: agenda item title, sometimes a subtitle or 'Before the vote:'); all heading lines are in meta.chapter_heading. Left out: the 111,208 chapters that have no SPEAKER turn at all (143,106 non-empty lines: agenda headings and procedural notes such as '(The sitting was closed at 23:40)'), which are not speeches; and the sentence-aligned X-English bitexts statmt.org distributes separately (europarl-v7.<x>-en.*), which are derived from these files. `date` = the sitting date from the file name; `parent_id` = the day file (ep-YY-MM-DD[-NNN..], the same in every language); `source_record_id` = '<lang>/<parallel_id>'. Translations: the language versions of a turn share meta.parallel_id = '<day file>/<CHAPTER ID>/<SPEAKER ID>' (+ '#k' for the k-th turn with the same ids in that chapter), the ids statmt.org's aligner matches across versions: 212,385 parallel ids, 36,495 of them in all 21 languages, 6,397 in one language only. meta also holds file_language (the folder), chapter_id, chapter_heading, speaker_id, speaker_name, speaker_language, speaker_affiliation (SPEAKER attributes as released; SPEAKER ID numbers the turns of a file, it is not a person id, and `author_id` is empty). speaker_language (LANGUAGE attribute) names the language the speech was given in and is set in 1,181,701 turns, almost always when it differs from the file's language (6,757 turns tag their own file language); where it is missing the version may be the original or a translation (tagging is uneven: e.g. txt/es/ tags almost no turns of 2000-2001). Most versions are therefore professional translations (human, but translated). One file with invalid UTF-8 is decoded with replacement characters (meta.utf8_errors_replaced, 74 rows).

*Licence:* statmt.org: "We are not aware of any copyright restrictions of the material" (cite Koehn 2005). The texts are the European Parliament's verbatim reports of proceedings, which the Parliament allows to be reused with acknowledgement of the source; licence id reuse-with-attribution. Credit: European Parliament; Europarl v7 (Koehn 2005).

### `govreport`

GovReport. Huang, Cao, Parulian, Ji & Wang (NAACL 2021).

Every report of the release: all 19,466 JSON files (crs/ 7,238, gao/ 12,228), in the release's train / valid / test splits (split_ids/<agency>_<split>.ids: CRS 6,514 / 362 / 362, GAO 11,005 / 612 / 611), one row per report, in split order, CRS before GAO, ids in the order of the id files. Nothing is filtered. `text` = the report body (CRS `reports`, GAO `report`: nested sections with section_title, paragraphs and subsections) written out in document order: each section's title as its own line (empty titles, such as the CRS root section, are skipped), then its paragraphs, then its subsections; title lines and paragraphs are separated by a blank line. All sections are kept, including GAO's top-level 'Letter' section with its own paragraphs (the paper's experiments dropped those paragraphs), appendices, agency comment letters, 'GAO Contact and Staff Acknowledgments' and 'Related GAO Products' lists. Paragraph and title text is as released: the PDF-extraction artefacts of the release are not repaired (double spaces where GAO lines were joined; UTF-8-read-as-Latin-1 mojibake such as 'â\x80\x94' for an em dash, in 568 CRS reports; split drop capitals such as 'T he' at a paragraph start, in 462 CRS reports; footnote numbers glued to words and section titles, 'Human Rights1'). `meta.summary` = the human-written summary as plain text in the same layout (CRS: the `summary` paragraphs; GAO: the `highlight` sections 'Why GAO Did This Study', 'What GAO Found', 'What GAO Recommends', ... with their titles); GAO rows also keep `highlight` (the structured summary as released) and `fastfact` (the release's shorter summary paragraphs, not used in the paper). Other meta fields as released: `agency` (CRS / GAO, from the folder), `id`, `released_date`, `published_date` (GAO only), `url` (GAO only). `title` = the report title; `date` = GAO published_date, else released_date (CRS has only released_date); `source_record_id` = '<agency>/<report id>'; `source_url` = the GAO product page (CRS reports carry no URL in the release). No author ids are released (reports are written by CRS and GAO staff).

*Licence:* CC BY 4.0 (dataset homepage gov-report-data.github.io: 'License CC BY 4.0'; the authors' HF copy launch/gov_report: cc-by-4.0). The reports themselves are works of the U.S. federal government (CRS and GAO), not subject to copyright in the U.S.; credit the GovReport authors.

### `hh_rlhf_harmless_base`

Anthropic HH-RLHF (helpfulness and harmlessness preference data, red-team attempts). Bai et al. (arXiv 2022), Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback; red-team attempts: Ganguli et al. (arXiv 2022), Red Teaming Language Models to Reduce Harms.

Every record of every file of the release, in five subsets, one per folder: `hh_rlhf_harmless_base`, `hh_rlhf_helpful_base`, `hh_rlhf_helpful_online`, `hh_rlhf_helpful_rejection_sampled` (preference pairs; splits `train` and `test` as released) and `hh_rlhf_red_team_attempts` (one file with no split, so `full`). Crowdworkers (MTurk and Upwork) wrote the Human turns; the Assistant turns are Anthropic's 52B models (context-distilled base models, rejection sampling against an early preference model, iterated online RLHF models; for red teaming also plain LMs and RLHF models of several sizes). The MIT licence covers the whole transcripts, so assistant turns are kept. Each transcript (`"\n\nHuman: ... \n\nAssistant: ..."`) is split at every `\n\nHuman:` / `\n\nAssistant:` marker into `messages` (user / assistant). As released, a few hundred transcripts have two turns of one role in a row or a turn that is empty, and some model turns contain the words `Human:` or `Assistant:` mid-line; these are kept as parsed. For a preference pair, `messages` is the `chosen` transcript and `meta.rejected` keeps the `rejected` transcript verbatim (the two usually differ only in the last assistant turn). For a red-team attempt, `messages` is `transcript` and `meta` keeps the other fields as released (`min_harmlessness_score_transcript`, `num_params`, `model_type`, `rating`, the red-teamer's `task_description`, `task_descripton_harmlessness_score` (sic), `red_team_member_id`, `is_upworker`, `tags`). `source_record_id` is `<folder>/<split>/<0-based line or item index>`. The release has no dates; `date` is the month of the paper that first describes the data (an upper bound): 2022-04 for the preference data (arXiv 2204.05862, 2022-04-12) and 2022-08 for the red-team attempts (arXiv 2209.07858; added to the GitHub release 2022-08-24).

*Licence:* MIT (dataset card; GitHub LICENSE "Copyright (c) 2022 Anthropic"). The card warns the data (especially harmlessness and red-team data) contain offensive or upsetting content, are intended for research (especially on making models less harmful), and are not meant for training dialogue agents.

### `hh_rlhf_helpful_base`

Anthropic HH-RLHF (helpfulness and harmlessness preference data, red-team attempts). Bai et al. (arXiv 2022), Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback; red-team attempts: Ganguli et al. (arXiv 2022), Red Teaming Language Models to Reduce Harms.

Every record of every file of the release, in five subsets, one per folder: `hh_rlhf_harmless_base`, `hh_rlhf_helpful_base`, `hh_rlhf_helpful_online`, `hh_rlhf_helpful_rejection_sampled` (preference pairs; splits `train` and `test` as released) and `hh_rlhf_red_team_attempts` (one file with no split, so `full`). Crowdworkers (MTurk and Upwork) wrote the Human turns; the Assistant turns are Anthropic's 52B models (context-distilled base models, rejection sampling against an early preference model, iterated online RLHF models; for red teaming also plain LMs and RLHF models of several sizes). The MIT licence covers the whole transcripts, so assistant turns are kept. Each transcript (`"\n\nHuman: ... \n\nAssistant: ..."`) is split at every `\n\nHuman:` / `\n\nAssistant:` marker into `messages` (user / assistant). As released, a few hundred transcripts have two turns of one role in a row or a turn that is empty, and some model turns contain the words `Human:` or `Assistant:` mid-line; these are kept as parsed. For a preference pair, `messages` is the `chosen` transcript and `meta.rejected` keeps the `rejected` transcript verbatim (the two usually differ only in the last assistant turn). For a red-team attempt, `messages` is `transcript` and `meta` keeps the other fields as released (`min_harmlessness_score_transcript`, `num_params`, `model_type`, `rating`, the red-teamer's `task_description`, `task_descripton_harmlessness_score` (sic), `red_team_member_id`, `is_upworker`, `tags`). `source_record_id` is `<folder>/<split>/<0-based line or item index>`. The release has no dates; `date` is the month of the paper that first describes the data (an upper bound): 2022-04 for the preference data (arXiv 2204.05862, 2022-04-12) and 2022-08 for the red-team attempts (arXiv 2209.07858; added to the GitHub release 2022-08-24).

*Licence:* MIT (dataset card; GitHub LICENSE "Copyright (c) 2022 Anthropic"). The card warns the data (especially harmlessness and red-team data) contain offensive or upsetting content, are intended for research (especially on making models less harmful), and are not meant for training dialogue agents.

### `hh_rlhf_helpful_online`

Anthropic HH-RLHF (helpfulness and harmlessness preference data, red-team attempts). Bai et al. (arXiv 2022), Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback; red-team attempts: Ganguli et al. (arXiv 2022), Red Teaming Language Models to Reduce Harms.

Every record of every file of the release, in five subsets, one per folder: `hh_rlhf_harmless_base`, `hh_rlhf_helpful_base`, `hh_rlhf_helpful_online`, `hh_rlhf_helpful_rejection_sampled` (preference pairs; splits `train` and `test` as released) and `hh_rlhf_red_team_attempts` (one file with no split, so `full`). Crowdworkers (MTurk and Upwork) wrote the Human turns; the Assistant turns are Anthropic's 52B models (context-distilled base models, rejection sampling against an early preference model, iterated online RLHF models; for red teaming also plain LMs and RLHF models of several sizes). The MIT licence covers the whole transcripts, so assistant turns are kept. Each transcript (`"\n\nHuman: ... \n\nAssistant: ..."`) is split at every `\n\nHuman:` / `\n\nAssistant:` marker into `messages` (user / assistant). As released, a few hundred transcripts have two turns of one role in a row or a turn that is empty, and some model turns contain the words `Human:` or `Assistant:` mid-line; these are kept as parsed. For a preference pair, `messages` is the `chosen` transcript and `meta.rejected` keeps the `rejected` transcript verbatim (the two usually differ only in the last assistant turn). For a red-team attempt, `messages` is `transcript` and `meta` keeps the other fields as released (`min_harmlessness_score_transcript`, `num_params`, `model_type`, `rating`, the red-teamer's `task_description`, `task_descripton_harmlessness_score` (sic), `red_team_member_id`, `is_upworker`, `tags`). `source_record_id` is `<folder>/<split>/<0-based line or item index>`. The release has no dates; `date` is the month of the paper that first describes the data (an upper bound): 2022-04 for the preference data (arXiv 2204.05862, 2022-04-12) and 2022-08 for the red-team attempts (arXiv 2209.07858; added to the GitHub release 2022-08-24).

*Licence:* MIT (dataset card; GitHub LICENSE "Copyright (c) 2022 Anthropic"). The card warns the data (especially harmlessness and red-team data) contain offensive or upsetting content, are intended for research (especially on making models less harmful), and are not meant for training dialogue agents.

### `hh_rlhf_helpful_rejection_sampled`

Anthropic HH-RLHF (helpfulness and harmlessness preference data, red-team attempts). Bai et al. (arXiv 2022), Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback; red-team attempts: Ganguli et al. (arXiv 2022), Red Teaming Language Models to Reduce Harms.

Every record of every file of the release, in five subsets, one per folder: `hh_rlhf_harmless_base`, `hh_rlhf_helpful_base`, `hh_rlhf_helpful_online`, `hh_rlhf_helpful_rejection_sampled` (preference pairs; splits `train` and `test` as released) and `hh_rlhf_red_team_attempts` (one file with no split, so `full`). Crowdworkers (MTurk and Upwork) wrote the Human turns; the Assistant turns are Anthropic's 52B models (context-distilled base models, rejection sampling against an early preference model, iterated online RLHF models; for red teaming also plain LMs and RLHF models of several sizes). The MIT licence covers the whole transcripts, so assistant turns are kept. Each transcript (`"\n\nHuman: ... \n\nAssistant: ..."`) is split at every `\n\nHuman:` / `\n\nAssistant:` marker into `messages` (user / assistant). As released, a few hundred transcripts have two turns of one role in a row or a turn that is empty, and some model turns contain the words `Human:` or `Assistant:` mid-line; these are kept as parsed. For a preference pair, `messages` is the `chosen` transcript and `meta.rejected` keeps the `rejected` transcript verbatim (the two usually differ only in the last assistant turn). For a red-team attempt, `messages` is `transcript` and `meta` keeps the other fields as released (`min_harmlessness_score_transcript`, `num_params`, `model_type`, `rating`, the red-teamer's `task_description`, `task_descripton_harmlessness_score` (sic), `red_team_member_id`, `is_upworker`, `tags`). `source_record_id` is `<folder>/<split>/<0-based line or item index>`. The release has no dates; `date` is the month of the paper that first describes the data (an upper bound): 2022-04 for the preference data (arXiv 2204.05862, 2022-04-12) and 2022-08 for the red-team attempts (arXiv 2209.07858; added to the GitHub release 2022-08-24).

*Licence:* MIT (dataset card; GitHub LICENSE "Copyright (c) 2022 Anthropic"). The card warns the data (especially harmlessness and red-team data) contain offensive or upsetting content, are intended for research (especially on making models less harmful), and are not meant for training dialogue agents.

### `hh_rlhf_red_team_attempts`

Anthropic HH-RLHF (helpfulness and harmlessness preference data, red-team attempts). Bai et al. (arXiv 2022), Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback; red-team attempts: Ganguli et al. (arXiv 2022), Red Teaming Language Models to Reduce Harms.

Every record of every file of the release, in five subsets, one per folder: `hh_rlhf_harmless_base`, `hh_rlhf_helpful_base`, `hh_rlhf_helpful_online`, `hh_rlhf_helpful_rejection_sampled` (preference pairs; splits `train` and `test` as released) and `hh_rlhf_red_team_attempts` (one file with no split, so `full`). Crowdworkers (MTurk and Upwork) wrote the Human turns; the Assistant turns are Anthropic's 52B models (context-distilled base models, rejection sampling against an early preference model, iterated online RLHF models; for red teaming also plain LMs and RLHF models of several sizes). The MIT licence covers the whole transcripts, so assistant turns are kept. Each transcript (`"\n\nHuman: ... \n\nAssistant: ..."`) is split at every `\n\nHuman:` / `\n\nAssistant:` marker into `messages` (user / assistant). As released, a few hundred transcripts have two turns of one role in a row or a turn that is empty, and some model turns contain the words `Human:` or `Assistant:` mid-line; these are kept as parsed. For a preference pair, `messages` is the `chosen` transcript and `meta.rejected` keeps the `rejected` transcript verbatim (the two usually differ only in the last assistant turn). For a red-team attempt, `messages` is `transcript` and `meta` keeps the other fields as released (`min_harmlessness_score_transcript`, `num_params`, `model_type`, `rating`, the red-teamer's `task_description`, `task_descripton_harmlessness_score` (sic), `red_team_member_id`, `is_upworker`, `tags`). `source_record_id` is `<folder>/<split>/<0-based line or item index>`. The release has no dates; `date` is the month of the paper that first describes the data (an upper bound): 2022-04 for the preference data (arXiv 2204.05862, 2022-04-12) and 2022-08 for the red-team attempts (arXiv 2209.07858; added to the GitHub release 2022-08-24).

*Licence:* MIT (dataset card; GitHub LICENSE "Copyright (c) 2022 Anthropic"). The card warns the data (especially harmlessness and red-team data) contain offensive or upsetting content, are intended for research (especially on making models less harmful), and are not meant for training dialogue agents.

### `ifbench`

IFBench: Generalizing Verifiable Instruction Following. Pyatkin, Malik, Graf, Ivison, Huang, Dasigi, Lambert & Hajishirzi (arXiv 2025).

All 300 rows of the release (one parquet; Hugging Face names its only split `train`, but the release is the IFBench test set, so the split here is `test`). `prompt` = the released `prompt`: a WildChat user prompt held out from WildChat's public release, with the constraint instruction(s) written into it. `meta.instruction_id_list` = the constraint ids (e.g. count:keywords_multiple) and `meta.kwargs` = their parameters, one dict per constraint; the parquet stores every possible parameter of every constraint with nulls, and only the non-null ones are kept (as in the repository's IFBench_test.jsonl). `source_record_id` = the release's `key`. No dates in the release.

*Licence:* ODC-BY-1.0 (dataset card: 'intended for research and educational use in accordance with Ai2's Responsible Use Guidelines'). The base prompts come from WildChat (ODC-BY).

### `issuebench`

IssueBench. Röttger, Hinck, Hofmann, Hackenburg, Pyatkin, Brahman & Hovy (TACL 2026).

The release has three configs, kept as three subsets. `issuebench` = config `prompts`, split `prompts_full`: all 2,490,576 prompts (3,916 templates x 212 issues x 3 framings). The release's other two prompt splits, `prompts_sample` (636,000, the set the paper evaluates) and `prompts_debug` (150), are exact subsets of `prompts_full` (same template, issue, framing and text, checked), so their rows are not repeated: `meta.in_prompts_sample` / `meta.in_prompts_debug` mark them. `prompt` = the released `prompt_text` (template with X replaced by the issue phrase); `meta` keeps template_id, template_text, topic_id, topic_polarity, topic_text; `source_record_id` = `<template_id>|<topic_id>|<topic_polarity>`. `issuebench_templates` = config `templates`, split `templates_full` (3,916; `templates_sample`, the 1,000 the paper uses, is an exact subset and is marked by `meta.in_templates_sample`): `prompt` = `user_prompt`, the logged first-turn prompt a chatbot user wrote (template_id prefix = the log it comes from: wildchat, lmsys, sharegpt, hhonline, prism); `meta` keeps the annotator's template (`annot1_template`, with placeholder X), label (minor edits / major edits / out of scope), annotator initial, notes and `n_duplicates`. `issuebench_issues` = config `issues`, split `issues` (396 issue clusters, including the 'UNASSIGNED CLUSTER' with topic_id -1; 212 are used in the prompts): these are not chatbot turns; `prompt` holds the neutral issue phrase (`topic_neutral`) and `meta` every other field as released (pro/con phrases, tags, cluster statistics, GPT-4o description, and `top_prompts` / `random_prompts`, example logged prompts of the cluster, as released strings). No dates per row (the logs span 2022-2024; issue phrases written by the authors before 2025-02). The authors' model completions (musashihinck/IssueBench_Completions) are a separate dataset and not included.

*Licence:* CC BY 4.0 (dataset card). The templates and logged prompts were taken from WildChat (ODC-BY), LMSYS-Chat-1M (LMSYS-Chat-1M Dataset License), ShareGPT90K (no stated licence), HH-RLHF (MIT) and PRISM; IssueBench redistributes them under CC BY 4.0. Users who need the upstream terms can filter on the template id prefix.

### `issuebench_issues`

IssueBench. Röttger, Hinck, Hofmann, Hackenburg, Pyatkin, Brahman & Hovy (TACL 2026).

The release has three configs, kept as three subsets. `issuebench` = config `prompts`, split `prompts_full`: all 2,490,576 prompts (3,916 templates x 212 issues x 3 framings). The release's other two prompt splits, `prompts_sample` (636,000, the set the paper evaluates) and `prompts_debug` (150), are exact subsets of `prompts_full` (same template, issue, framing and text, checked), so their rows are not repeated: `meta.in_prompts_sample` / `meta.in_prompts_debug` mark them. `prompt` = the released `prompt_text` (template with X replaced by the issue phrase); `meta` keeps template_id, template_text, topic_id, topic_polarity, topic_text; `source_record_id` = `<template_id>|<topic_id>|<topic_polarity>`. `issuebench_templates` = config `templates`, split `templates_full` (3,916; `templates_sample`, the 1,000 the paper uses, is an exact subset and is marked by `meta.in_templates_sample`): `prompt` = `user_prompt`, the logged first-turn prompt a chatbot user wrote (template_id prefix = the log it comes from: wildchat, lmsys, sharegpt, hhonline, prism); `meta` keeps the annotator's template (`annot1_template`, with placeholder X), label (minor edits / major edits / out of scope), annotator initial, notes and `n_duplicates`. `issuebench_issues` = config `issues`, split `issues` (396 issue clusters, including the 'UNASSIGNED CLUSTER' with topic_id -1; 212 are used in the prompts): these are not chatbot turns; `prompt` holds the neutral issue phrase (`topic_neutral`) and `meta` every other field as released (pro/con phrases, tags, cluster statistics, GPT-4o description, and `top_prompts` / `random_prompts`, example logged prompts of the cluster, as released strings). No dates per row (the logs span 2022-2024; issue phrases written by the authors before 2025-02). The authors' model completions (musashihinck/IssueBench_Completions) are a separate dataset and not included.

*Licence:* CC BY 4.0 (dataset card). The templates and logged prompts were taken from WildChat (ODC-BY), LMSYS-Chat-1M (LMSYS-Chat-1M Dataset License), ShareGPT90K (no stated licence), HH-RLHF (MIT) and PRISM; IssueBench redistributes them under CC BY 4.0. Users who need the upstream terms can filter on the template id prefix.

### `issuebench_templates`

IssueBench. Röttger, Hinck, Hofmann, Hackenburg, Pyatkin, Brahman & Hovy (TACL 2026).

The release has three configs, kept as three subsets. `issuebench` = config `prompts`, split `prompts_full`: all 2,490,576 prompts (3,916 templates x 212 issues x 3 framings). The release's other two prompt splits, `prompts_sample` (636,000, the set the paper evaluates) and `prompts_debug` (150), are exact subsets of `prompts_full` (same template, issue, framing and text, checked), so their rows are not repeated: `meta.in_prompts_sample` / `meta.in_prompts_debug` mark them. `prompt` = the released `prompt_text` (template with X replaced by the issue phrase); `meta` keeps template_id, template_text, topic_id, topic_polarity, topic_text; `source_record_id` = `<template_id>|<topic_id>|<topic_polarity>`. `issuebench_templates` = config `templates`, split `templates_full` (3,916; `templates_sample`, the 1,000 the paper uses, is an exact subset and is marked by `meta.in_templates_sample`): `prompt` = `user_prompt`, the logged first-turn prompt a chatbot user wrote (template_id prefix = the log it comes from: wildchat, lmsys, sharegpt, hhonline, prism); `meta` keeps the annotator's template (`annot1_template`, with placeholder X), label (minor edits / major edits / out of scope), annotator initial, notes and `n_duplicates`. `issuebench_issues` = config `issues`, split `issues` (396 issue clusters, including the 'UNASSIGNED CLUSTER' with topic_id -1; 212 are used in the prompts): these are not chatbot turns; `prompt` holds the neutral issue phrase (`topic_neutral`) and `meta` every other field as released (pro/con phrases, tags, cluster statistics, GPT-4o description, and `top_prompts` / `random_prompts`, example logged prompts of the cluster, as released strings). No dates per row (the logs span 2022-2024; issue phrases written by the authors before 2025-02). The authors' model completions (musashihinck/IssueBench_Completions) are a separate dataset and not included.

*Licence:* CC BY 4.0 (dataset card). The templates and logged prompts were taken from WildChat (ODC-BY), LMSYS-Chat-1M (LMSYS-Chat-1M Dataset License), ShareGPT90K (no stated licence), HH-RLHF (MIT) and PRISM; IssueBench redistributes them under CC BY 4.0. Users who need the upstream terms can filter on the template id prefix.

### `iterater_full_doc`

IteraTeR (document level: IteraTeR-FULL and IteraTeR-HUMAN). Du, Raheja, Kumar, Kim, Lopez & Kang (ACL 2022).

Every document-level revision pair of both releases, all splits: IteraTeR-FULL train 29,848 / dev 856 / test 927 pairs (subset iterater_full_doc) and IteraTeR-HUMAN train 481 / dev 27 / test 51 (subset iterater_human_doc; an annotated sample of documents that are mostly also in FULL, so texts repeat across the two subsets). The sentence-level configs are left out (they cut the same pairs into single sentences). Each pair gives two rows, one per document version: the `before_revision` text (version = revision_depth - 1) and the `after_revision` text (version = revision_depth), in file order. Versions are not merged across pairs: in the release the 'after' text of depth t and the 'before' text of depth t+1 of the same document are mostly not identical (FULL: 378 of 6,100 such links identical, often only whitespace differs for arXiv, while many Wikipedia/Wikinews links differ substantially), so identical versions can repeat; `text_hash` finds them. `text` is the version as released (the authors' plain-text rendering: arXiv abstracts with latexdiff tokenisation, wiki markup removed by their parser, URLs replaced by 'URL'); nothing is filtered. `parent_id` = '<domain>/<doc_id>' (the document: arXiv id or MediaWiki page id). `meta`: `doc_id`, `revision_depth` (as released: text in FULL, number in HUMAN), `version`, `side` (before / after), `pair` (source_record_id of the other row of the pair), `domain` (as released; HUMAN only, missing for 38 HUMAN pairs), `domain_derived` and `domain_evidence`; the 'after' row also holds the pair's annotations as released: `edit_actions` (type A/D/R, before/after strings, start/end character offsets into the before text; HUMAN adds `major_intent` and the annotators' `raw_intents`) and `sents_char_pos` (sentence start offsets in the before text). Domain (only used for licence and genre): the released `domain`, else the arXiv id pattern (arXiv), else the MediaWiki page id looked up on en.wikipedia and en.wikinews (page ids are per wiki and collide: FULL train has 22 doc_id + depth keys that occur twice, for two different documents): the wiki whose current page title has more of its words (>2 letters) in the first 3,000 characters of the pair's texts, with at least half of them; if neither title qualifies, the only wiki that has a page with that id; else unresolved ('wiki_or_news'). Checked on the 359 HUMAN pairs whose released domain is wiki or news (domain withheld): 332 right, 0 wrong, 27 unresolved (26 of them Wikinews, whose texts often lost the linked names of the title). FULL result: arxiv 11,443 pairs, wiki 10,473, news 7,527, unresolved 2,188. `date` = '2021' for every row, an upper bound: the authors collected all versions in July 2021 (paper appendix A), the release has no per-version dates. No titles or author ids are released. `source_record_id` = '<config>/<split>/<0-based line number in the split file>/<before|after>' (config full_doc or human_doc, so ids stay distinct across the two subsets).

*Licence:* The release is Apache-2.0 (HF dataset cards and GitHub LICENSE), but the texts keep their own terms, and the strictest reading is used per row: arXiv abstracts (arXiv metadata is CC0) -> apache-2.0; Wikipedia text -> cc-by-sa-3.0 (Wikipedia's text licence before June 2023; share-alike); Wikinews text -> cc-by-2.5 (Wikinews' licence); a non-arXiv document whose wiki is unresolved -> cc-by-sa-3.0. All allow commercial use with attribution.

### `iterater_human_doc`

IteraTeR (document level: IteraTeR-FULL and IteraTeR-HUMAN). Du, Raheja, Kumar, Kim, Lopez & Kang (ACL 2022).

Every document-level revision pair of both releases, all splits: IteraTeR-FULL train 29,848 / dev 856 / test 927 pairs (subset iterater_full_doc) and IteraTeR-HUMAN train 481 / dev 27 / test 51 (subset iterater_human_doc; an annotated sample of documents that are mostly also in FULL, so texts repeat across the two subsets). The sentence-level configs are left out (they cut the same pairs into single sentences). Each pair gives two rows, one per document version: the `before_revision` text (version = revision_depth - 1) and the `after_revision` text (version = revision_depth), in file order. Versions are not merged across pairs: in the release the 'after' text of depth t and the 'before' text of depth t+1 of the same document are mostly not identical (FULL: 378 of 6,100 such links identical, often only whitespace differs for arXiv, while many Wikipedia/Wikinews links differ substantially), so identical versions can repeat; `text_hash` finds them. `text` is the version as released (the authors' plain-text rendering: arXiv abstracts with latexdiff tokenisation, wiki markup removed by their parser, URLs replaced by 'URL'); nothing is filtered. `parent_id` = '<domain>/<doc_id>' (the document: arXiv id or MediaWiki page id). `meta`: `doc_id`, `revision_depth` (as released: text in FULL, number in HUMAN), `version`, `side` (before / after), `pair` (source_record_id of the other row of the pair), `domain` (as released; HUMAN only, missing for 38 HUMAN pairs), `domain_derived` and `domain_evidence`; the 'after' row also holds the pair's annotations as released: `edit_actions` (type A/D/R, before/after strings, start/end character offsets into the before text; HUMAN adds `major_intent` and the annotators' `raw_intents`) and `sents_char_pos` (sentence start offsets in the before text). Domain (only used for licence and genre): the released `domain`, else the arXiv id pattern (arXiv), else the MediaWiki page id looked up on en.wikipedia and en.wikinews (page ids are per wiki and collide: FULL train has 22 doc_id + depth keys that occur twice, for two different documents): the wiki whose current page title has more of its words (>2 letters) in the first 3,000 characters of the pair's texts, with at least half of them; if neither title qualifies, the only wiki that has a page with that id; else unresolved ('wiki_or_news'). Checked on the 359 HUMAN pairs whose released domain is wiki or news (domain withheld): 332 right, 0 wrong, 27 unresolved (26 of them Wikinews, whose texts often lost the linked names of the title). FULL result: arxiv 11,443 pairs, wiki 10,473, news 7,527, unresolved 2,188. `date` = '2021' for every row, an upper bound: the authors collected all versions in July 2021 (paper appendix A), the release has no per-version dates. No titles or author ids are released. `source_record_id` = '<config>/<split>/<0-based line number in the split file>/<before|after>' (config full_doc or human_doc, so ids stay distinct across the two subsets).

*Licence:* The release is Apache-2.0 (HF dataset cards and GitHub LICENSE), but the texts keep their own terms, and the strictest reading is used per row: arXiv abstracts (arXiv metadata is CC0) -> apache-2.0; Wikipedia text -> cc-by-sa-3.0 (Wikipedia's text licence before June 2023; share-alike); Wikinews text -> cc-by-2.5 (Wikinews' licence); a non-arXiv document whose wiki is unresolved -> cc-by-sa-3.0. All allow commercial use with attribution.

### `kilt`

KILT knowledge source (English Wikipedia, 2019-08-01). Petroni, Piktus, Fan, Lewis, Yazdani, De Cao, Thorne, Jernite, Karpukhin, Maillard, Plachouras, Rocktäschel & Riedel (NAACL 2021).

Every line of the knowledge source (5,903,530 pages of the 2019-08-01 English Wikipedia snapshot that KILT preprocessed), one row per page, in file order; split `full` (the knowledge source has no splits). Nothing is sampled or filtered. 14,177 pages hold only their title line (KILT kept no paragraph for them, e.g. 'Meanings of minor planet names: 88001–89000', a page that is one table); they have no text and are counted by the builder as dropped_empty, leaving 5,889,353 rows. `text` is built from the release's paragraph list: text[0] (the title line) is left out when it is the page title (after HTML-unescaping, e.g. 'T &amp; T Supermarket'), otherwise kept as the first paragraph (meta.title_line_kept); every other paragraph is kept, in order, with its trailing newline removed; a 'Section::::History.:Typographic variants.' marker becomes a heading line holding the last section name with KILT's added final '.' removed ('Typographic variants'); a list item 'BULLET::::- x' or 'BULLET::::1. x' becomes '- x' or '1. x' (the marker removed). Blocks are separated by a blank line, except consecutive list items, which are separated by one newline. All sections are kept, including See also / References / External links (KILT already removed tables, infoboxes, references markup and templates). `title` = wikipedia_title; `source_record_id` = wikipedia_id (the page id); `date` = history.timestamp (the timestamp of the page revision KILT used), or the snapshot date 2019-08-01 (an upper bound; meta.date_is_snapshot) for the pages KILT found no revision for (history.revid null); `source_url` = history.url (that revision's oldid URL; none when revid is null, where KILT's URL ends in 'oldid=None'); `genre` = 'encyclopedia article'. `meta` holds wikipedia_id, _id, n_anchors (the number of hyperlinks KILT records for the page; the anchor list itself is left out), categories (as released, comma-joined), history (revid, parentid, pageid, timestamp, pre_dump, url), wikidata_info (as released) and sections ([depth, heading] for every section marker, depth 1 = top-level section). Overlap: these are largely the same articles as the `wikipedia` source (2022-05-01 dump) in an older revision, so near-duplicate texts across the two sources are expected. The KILT task files (train/dev/test of the 11 KILT datasets) are not part of the knowledge source and are not used.

*Licence:* The KILT release says 'KILT is MIT licensed'; the page texts are English Wikipedia content, which Wikipedia licensed under CC BY-SA 3.0 (and GFDL) at the 2019 snapshot and which a third party cannot relicense. The stricter reading is used: every row is CC BY-SA 3.0. Attribution: the page history (history.url); share-alike applies.

### `kiwi`

KIWI: A Dataset of Knowledge-Intensive Writing Instructions for Answering Research Questions. Xu, Lo, Soldaini, Kuehl, Choi & Wadden (Findings of ACL 2024).

Every session of the release (one file, 234 sessions = 78 research questions x 3 models: gpt-4, gpt-3.5-turbo, llama-2; 1,260 instruction turns; no split, so the split is `full`). Each session is one conversation: user = `original_question` (written by the researcher), assistant = `initial_answer`, then for every turn of `interaction` user = `instruction` (the researcher's revision request) and assistant = `answer_1` (the model's revised answer). `model` = `model_name`. `meta` keeps annotator_id, `initial_docs` (the retrieved paper passages behind the initial answer, cited as [n]) and, per turn, the fields not in `messages`: `answer_2` (the researcher's own edit of the answer, when they made one: 340 turns), `rating` (good/neutral/bad), `comment` and `docs`. One instruction is empty in the release and stays an empty user turn. The release has no ids; `source_record_id` = `session_<line index>`. Sessions were collected July-December 2023 (paper Sec. 3.2).

*Licence:* CC BY-SA 4.0 (dataset card, Licensing Information; the card's YAML tag is the generic 'cc'). It covers the whole sessions, including the model answers and the retrieved passages.

### `klicke`

KLiCKe: Keystroke Logging in Compositions for Knowledge Evaluation. Tian, Crossley & Van Waes (Journal of Writing Research 2025).

Every file of the official release except the two Word documents. (1) `klicke` (documents, split `full`; the release has no splits): all 4,992 essays, one row per WritingTask/texts/<ID>.txt, in ID order. `text` = the .txt file (UTF-8; 11 files already contain U+FFFD replacement characters where the release lost a character); `prompt` = the full SAT-based writing prompt of the essay's topic, taken from WritingTask/writing_prompts.docx (the paragraph after each topic cell: APPEARANCE, COMPETITION, HAPPINESS, MATERIALISM); `source_record_id` = `author_id` = the release's 8-digit anonymised ID (one essay per writer); `date` = 2022 (README and paper: collected January-November 2022 on MTurk, 'before the release of ChatGPT'; no per-essay date is released, so the builder's pre_chatgpt is null for the year). `meta` = `Prompt` (topic) and `Score` (holistic score 1-6) from holistic_scores.csv, `demographics` = the writer's demographic_info.csv row as released (all 16 columns incl. the unnamed row index; Gender, Age, Nationality, Ethnicity, Multiple_Ethnicity, Highest_Degree, Handedness, Nativeness, Language, Exposure_Age, Exposure_Length, Length_of_Stay, Writing_Interest, Writing_Strength), and `holistic_scores_text` = the essay's `Text` cell in holistic_scores.csv only for the 11 essays where it differs from the .txt file (exactly the 11 whose .txt has U+FFFD; the CSV cell has the original character, e.g. a curly quote or no-break space); for the other 4,981 the cell equals the txt file and is not repeated (holistic_scores.csv is UTF-8 with stray cp1252 bytes, decoded as for the logs below). (2) `klicke_keystrokes` (logs): 9,984 rows, the two keystroke logs of every essay exactly as released (keystrokelogs/csv/<ID>.csv, format `csv`, the authors' JavaScript logger output; keystrokelogs/idfx/<ID>.idfx, format `idfx`, the same log converted for Inputlog); `source_record_id` = '<ID>/csv' or '<ID>/idfx', `parent_id` = the essay's row id 'klicke/<ID>'. (3) `klicke_typing_tests` (logs): 69,840 rows, every file of TypingTests.zip: the Inputlog copy tasks (Van Waes et al. 2019) each writer took, as CSV and IDFX logs, named <ID>_<test> with tests TYPING1, TYPING2, TYPING3_1..TYPING3_4, TYPING4 (4,985-4,991 writers per test); `source_record_id` = '<ID>_<test>/<format>', `parent_id` = 'klicke/<ID>', meta.test = the test name. Copy-task text is given text retyped, not the writer's own writing, so it stays a log and is not a document. (4) `klicke_vocabulary_test` (logs): 4,971 rows, every file of VocabularyKnowledgeTest.zip (the writer's LexTALE responses: item, word, response, key), format `csv`, `source_record_id` = '<ID>/vocab', `parent_id` = 'klicke/<ID>'. Log files are decoded as UTF-8; 21 essay CSV logs and 3 typing-test CSV logs contain bytes that are not valid UTF-8 (single cp1252 bytes such as 0x92): those bytes alone are decoded as cp1252 (latin-1 for the five bytes cp1252 leaves undefined) and meta.encoding says so; all other logs and all texts are plain UTF-8. Logs keep their CRLF line ends and every byte as released apart from the worker-ID replacement below. Worker IDs are replaced (owner decision 2026-10-13): the release keeps each writer's Amazon Mechanical Turk worker ID (21 values do not have the usual 'A' + 11-13 character form: 5 short numbers, longer codes, one e-mail address) in the Participant and __LogFileName fields of the writer's essay and typing-test IDFX logs, and some writers pasted or typed their ID into the essay (1 text) or during the tasks (it then appears in the CSV and IDFX logs). The writer's ID is read from the Participant field of their essay IDFX log (one distinct value per writer; the typing tests use '<ID>_<test>') and replaced by a code: 'W' + the writer's 4-digit rank in sorted release-ID order (10010800 -> W0001), padded with '_' to the length of the ID so that character positions in the logs stay valid, e.g. 'W0001_________' and 'W0001__________TYPING1'. The same code is used for that writer in every subset. IDs of 8 or more characters with a letter (all but 5) are replaced wherever they occur in the writer's essay text, holistic_scores_text, keystroke, typing-test and vocabulary-test logs, and so is any piece of 8 or more characters of such an ID (with a letter) found inside a run of capitals and digits (e.g. the part of a pasted ID that a writer then cut, which the logs record as removed text), with the code padded to the piece's length; the 5 short numeric Participant values (e.g. a 4-digit number) are replaced only in the two IDFX fields, as elsewhere they are ordinary numbers. No mapping is stored. meta.worker_id_replaced counts the replacements in a row. These are the only changes to texts and logs. A scan of the built files (case-insensitive) finds none of the original IDs and no piece of 8 or more of their characters. An ID typed key by key leaves one character per log event; such sequences are not reassembled. Not changed: an MTurk requester ID (the public ID of a requester account, not a worker) in a turkerview.com URL that one writer (83107247) pasted into the essay. Left out: WritingTask/writing_prompts.docx (its prompt texts are in `prompt`) and WritingTask/holistic_rating_form.docx (the raters' rubric), which are not records. Nothing is sampled or filtered.

*Licence:* MIT (LICENSE in the official GitHub repo, (c) 2025 terryyutian; the README links the Drive data, which state no other licence). The JoWR paper itself is CC BY-NC-ND 3.0. Privacy: the official release keeps the writers' Amazon Mechanical Turk worker IDs in the IDFX logs (and a few texts and CSV logs); here every one is replaced by an anonymous code (W0001...), see processing. MIT permits the modification.

### `klicke_keystrokes`

KLiCKe: Keystroke Logging in Compositions for Knowledge Evaluation. Tian, Crossley & Van Waes (Journal of Writing Research 2025).

Every file of the official release except the two Word documents. (1) `klicke` (documents, split `full`; the release has no splits): all 4,992 essays, one row per WritingTask/texts/<ID>.txt, in ID order. `text` = the .txt file (UTF-8; 11 files already contain U+FFFD replacement characters where the release lost a character); `prompt` = the full SAT-based writing prompt of the essay's topic, taken from WritingTask/writing_prompts.docx (the paragraph after each topic cell: APPEARANCE, COMPETITION, HAPPINESS, MATERIALISM); `source_record_id` = `author_id` = the release's 8-digit anonymised ID (one essay per writer); `date` = 2022 (README and paper: collected January-November 2022 on MTurk, 'before the release of ChatGPT'; no per-essay date is released, so the builder's pre_chatgpt is null for the year). `meta` = `Prompt` (topic) and `Score` (holistic score 1-6) from holistic_scores.csv, `demographics` = the writer's demographic_info.csv row as released (all 16 columns incl. the unnamed row index; Gender, Age, Nationality, Ethnicity, Multiple_Ethnicity, Highest_Degree, Handedness, Nativeness, Language, Exposure_Age, Exposure_Length, Length_of_Stay, Writing_Interest, Writing_Strength), and `holistic_scores_text` = the essay's `Text` cell in holistic_scores.csv only for the 11 essays where it differs from the .txt file (exactly the 11 whose .txt has U+FFFD; the CSV cell has the original character, e.g. a curly quote or no-break space); for the other 4,981 the cell equals the txt file and is not repeated (holistic_scores.csv is UTF-8 with stray cp1252 bytes, decoded as for the logs below). (2) `klicke_keystrokes` (logs): 9,984 rows, the two keystroke logs of every essay exactly as released (keystrokelogs/csv/<ID>.csv, format `csv`, the authors' JavaScript logger output; keystrokelogs/idfx/<ID>.idfx, format `idfx`, the same log converted for Inputlog); `source_record_id` = '<ID>/csv' or '<ID>/idfx', `parent_id` = the essay's row id 'klicke/<ID>'. (3) `klicke_typing_tests` (logs): 69,840 rows, every file of TypingTests.zip: the Inputlog copy tasks (Van Waes et al. 2019) each writer took, as CSV and IDFX logs, named <ID>_<test> with tests TYPING1, TYPING2, TYPING3_1..TYPING3_4, TYPING4 (4,985-4,991 writers per test); `source_record_id` = '<ID>_<test>/<format>', `parent_id` = 'klicke/<ID>', meta.test = the test name. Copy-task text is given text retyped, not the writer's own writing, so it stays a log and is not a document. (4) `klicke_vocabulary_test` (logs): 4,971 rows, every file of VocabularyKnowledgeTest.zip (the writer's LexTALE responses: item, word, response, key), format `csv`, `source_record_id` = '<ID>/vocab', `parent_id` = 'klicke/<ID>'. Log files are decoded as UTF-8; 21 essay CSV logs and 3 typing-test CSV logs contain bytes that are not valid UTF-8 (single cp1252 bytes such as 0x92): those bytes alone are decoded as cp1252 (latin-1 for the five bytes cp1252 leaves undefined) and meta.encoding says so; all other logs and all texts are plain UTF-8. Logs keep their CRLF line ends and every byte as released apart from the worker-ID replacement below. Worker IDs are replaced (owner decision 2026-10-13): the release keeps each writer's Amazon Mechanical Turk worker ID (21 values do not have the usual 'A' + 11-13 character form: 5 short numbers, longer codes, one e-mail address) in the Participant and __LogFileName fields of the writer's essay and typing-test IDFX logs, and some writers pasted or typed their ID into the essay (1 text) or during the tasks (it then appears in the CSV and IDFX logs). The writer's ID is read from the Participant field of their essay IDFX log (one distinct value per writer; the typing tests use '<ID>_<test>') and replaced by a code: 'W' + the writer's 4-digit rank in sorted release-ID order (10010800 -> W0001), padded with '_' to the length of the ID so that character positions in the logs stay valid, e.g. 'W0001_________' and 'W0001__________TYPING1'. The same code is used for that writer in every subset. IDs of 8 or more characters with a letter (all but 5) are replaced wherever they occur in the writer's essay text, holistic_scores_text, keystroke, typing-test and vocabulary-test logs, and so is any piece of 8 or more characters of such an ID (with a letter) found inside a run of capitals and digits (e.g. the part of a pasted ID that a writer then cut, which the logs record as removed text), with the code padded to the piece's length; the 5 short numeric Participant values (e.g. a 4-digit number) are replaced only in the two IDFX fields, as elsewhere they are ordinary numbers. No mapping is stored. meta.worker_id_replaced counts the replacements in a row. These are the only changes to texts and logs. A scan of the built files (case-insensitive) finds none of the original IDs and no piece of 8 or more of their characters. An ID typed key by key leaves one character per log event; such sequences are not reassembled. Not changed: an MTurk requester ID (the public ID of a requester account, not a worker) in a turkerview.com URL that one writer (83107247) pasted into the essay. Left out: WritingTask/writing_prompts.docx (its prompt texts are in `prompt`) and WritingTask/holistic_rating_form.docx (the raters' rubric), which are not records. Nothing is sampled or filtered.

*Licence:* MIT (LICENSE in the official GitHub repo, (c) 2025 terryyutian; the README links the Drive data, which state no other licence). The JoWR paper itself is CC BY-NC-ND 3.0. Privacy: the official release keeps the writers' Amazon Mechanical Turk worker IDs in the IDFX logs (and a few texts and CSV logs); here every one is replaced by an anonymous code (W0001...), see processing. MIT permits the modification.

### `klicke_typing_tests`

KLiCKe: Keystroke Logging in Compositions for Knowledge Evaluation. Tian, Crossley & Van Waes (Journal of Writing Research 2025).

Every file of the official release except the two Word documents. (1) `klicke` (documents, split `full`; the release has no splits): all 4,992 essays, one row per WritingTask/texts/<ID>.txt, in ID order. `text` = the .txt file (UTF-8; 11 files already contain U+FFFD replacement characters where the release lost a character); `prompt` = the full SAT-based writing prompt of the essay's topic, taken from WritingTask/writing_prompts.docx (the paragraph after each topic cell: APPEARANCE, COMPETITION, HAPPINESS, MATERIALISM); `source_record_id` = `author_id` = the release's 8-digit anonymised ID (one essay per writer); `date` = 2022 (README and paper: collected January-November 2022 on MTurk, 'before the release of ChatGPT'; no per-essay date is released, so the builder's pre_chatgpt is null for the year). `meta` = `Prompt` (topic) and `Score` (holistic score 1-6) from holistic_scores.csv, `demographics` = the writer's demographic_info.csv row as released (all 16 columns incl. the unnamed row index; Gender, Age, Nationality, Ethnicity, Multiple_Ethnicity, Highest_Degree, Handedness, Nativeness, Language, Exposure_Age, Exposure_Length, Length_of_Stay, Writing_Interest, Writing_Strength), and `holistic_scores_text` = the essay's `Text` cell in holistic_scores.csv only for the 11 essays where it differs from the .txt file (exactly the 11 whose .txt has U+FFFD; the CSV cell has the original character, e.g. a curly quote or no-break space); for the other 4,981 the cell equals the txt file and is not repeated (holistic_scores.csv is UTF-8 with stray cp1252 bytes, decoded as for the logs below). (2) `klicke_keystrokes` (logs): 9,984 rows, the two keystroke logs of every essay exactly as released (keystrokelogs/csv/<ID>.csv, format `csv`, the authors' JavaScript logger output; keystrokelogs/idfx/<ID>.idfx, format `idfx`, the same log converted for Inputlog); `source_record_id` = '<ID>/csv' or '<ID>/idfx', `parent_id` = the essay's row id 'klicke/<ID>'. (3) `klicke_typing_tests` (logs): 69,840 rows, every file of TypingTests.zip: the Inputlog copy tasks (Van Waes et al. 2019) each writer took, as CSV and IDFX logs, named <ID>_<test> with tests TYPING1, TYPING2, TYPING3_1..TYPING3_4, TYPING4 (4,985-4,991 writers per test); `source_record_id` = '<ID>_<test>/<format>', `parent_id` = 'klicke/<ID>', meta.test = the test name. Copy-task text is given text retyped, not the writer's own writing, so it stays a log and is not a document. (4) `klicke_vocabulary_test` (logs): 4,971 rows, every file of VocabularyKnowledgeTest.zip (the writer's LexTALE responses: item, word, response, key), format `csv`, `source_record_id` = '<ID>/vocab', `parent_id` = 'klicke/<ID>'. Log files are decoded as UTF-8; 21 essay CSV logs and 3 typing-test CSV logs contain bytes that are not valid UTF-8 (single cp1252 bytes such as 0x92): those bytes alone are decoded as cp1252 (latin-1 for the five bytes cp1252 leaves undefined) and meta.encoding says so; all other logs and all texts are plain UTF-8. Logs keep their CRLF line ends and every byte as released apart from the worker-ID replacement below. Worker IDs are replaced (owner decision 2026-10-13): the release keeps each writer's Amazon Mechanical Turk worker ID (21 values do not have the usual 'A' + 11-13 character form: 5 short numbers, longer codes, one e-mail address) in the Participant and __LogFileName fields of the writer's essay and typing-test IDFX logs, and some writers pasted or typed their ID into the essay (1 text) or during the tasks (it then appears in the CSV and IDFX logs). The writer's ID is read from the Participant field of their essay IDFX log (one distinct value per writer; the typing tests use '<ID>_<test>') and replaced by a code: 'W' + the writer's 4-digit rank in sorted release-ID order (10010800 -> W0001), padded with '_' to the length of the ID so that character positions in the logs stay valid, e.g. 'W0001_________' and 'W0001__________TYPING1'. The same code is used for that writer in every subset. IDs of 8 or more characters with a letter (all but 5) are replaced wherever they occur in the writer's essay text, holistic_scores_text, keystroke, typing-test and vocabulary-test logs, and so is any piece of 8 or more characters of such an ID (with a letter) found inside a run of capitals and digits (e.g. the part of a pasted ID that a writer then cut, which the logs record as removed text), with the code padded to the piece's length; the 5 short numeric Participant values (e.g. a 4-digit number) are replaced only in the two IDFX fields, as elsewhere they are ordinary numbers. No mapping is stored. meta.worker_id_replaced counts the replacements in a row. These are the only changes to texts and logs. A scan of the built files (case-insensitive) finds none of the original IDs and no piece of 8 or more of their characters. An ID typed key by key leaves one character per log event; such sequences are not reassembled. Not changed: an MTurk requester ID (the public ID of a requester account, not a worker) in a turkerview.com URL that one writer (83107247) pasted into the essay. Left out: WritingTask/writing_prompts.docx (its prompt texts are in `prompt`) and WritingTask/holistic_rating_form.docx (the raters' rubric), which are not records. Nothing is sampled or filtered.

*Licence:* MIT (LICENSE in the official GitHub repo, (c) 2025 terryyutian; the README links the Drive data, which state no other licence). The JoWR paper itself is CC BY-NC-ND 3.0. Privacy: the official release keeps the writers' Amazon Mechanical Turk worker IDs in the IDFX logs (and a few texts and CSV logs); here every one is replaced by an anonymous code (W0001...), see processing. MIT permits the modification.

### `klicke_vocabulary_test`

KLiCKe: Keystroke Logging in Compositions for Knowledge Evaluation. Tian, Crossley & Van Waes (Journal of Writing Research 2025).

Every file of the official release except the two Word documents. (1) `klicke` (documents, split `full`; the release has no splits): all 4,992 essays, one row per WritingTask/texts/<ID>.txt, in ID order. `text` = the .txt file (UTF-8; 11 files already contain U+FFFD replacement characters where the release lost a character); `prompt` = the full SAT-based writing prompt of the essay's topic, taken from WritingTask/writing_prompts.docx (the paragraph after each topic cell: APPEARANCE, COMPETITION, HAPPINESS, MATERIALISM); `source_record_id` = `author_id` = the release's 8-digit anonymised ID (one essay per writer); `date` = 2022 (README and paper: collected January-November 2022 on MTurk, 'before the release of ChatGPT'; no per-essay date is released, so the builder's pre_chatgpt is null for the year). `meta` = `Prompt` (topic) and `Score` (holistic score 1-6) from holistic_scores.csv, `demographics` = the writer's demographic_info.csv row as released (all 16 columns incl. the unnamed row index; Gender, Age, Nationality, Ethnicity, Multiple_Ethnicity, Highest_Degree, Handedness, Nativeness, Language, Exposure_Age, Exposure_Length, Length_of_Stay, Writing_Interest, Writing_Strength), and `holistic_scores_text` = the essay's `Text` cell in holistic_scores.csv only for the 11 essays where it differs from the .txt file (exactly the 11 whose .txt has U+FFFD; the CSV cell has the original character, e.g. a curly quote or no-break space); for the other 4,981 the cell equals the txt file and is not repeated (holistic_scores.csv is UTF-8 with stray cp1252 bytes, decoded as for the logs below). (2) `klicke_keystrokes` (logs): 9,984 rows, the two keystroke logs of every essay exactly as released (keystrokelogs/csv/<ID>.csv, format `csv`, the authors' JavaScript logger output; keystrokelogs/idfx/<ID>.idfx, format `idfx`, the same log converted for Inputlog); `source_record_id` = '<ID>/csv' or '<ID>/idfx', `parent_id` = the essay's row id 'klicke/<ID>'. (3) `klicke_typing_tests` (logs): 69,840 rows, every file of TypingTests.zip: the Inputlog copy tasks (Van Waes et al. 2019) each writer took, as CSV and IDFX logs, named <ID>_<test> with tests TYPING1, TYPING2, TYPING3_1..TYPING3_4, TYPING4 (4,985-4,991 writers per test); `source_record_id` = '<ID>_<test>/<format>', `parent_id` = 'klicke/<ID>', meta.test = the test name. Copy-task text is given text retyped, not the writer's own writing, so it stays a log and is not a document. (4) `klicke_vocabulary_test` (logs): 4,971 rows, every file of VocabularyKnowledgeTest.zip (the writer's LexTALE responses: item, word, response, key), format `csv`, `source_record_id` = '<ID>/vocab', `parent_id` = 'klicke/<ID>'. Log files are decoded as UTF-8; 21 essay CSV logs and 3 typing-test CSV logs contain bytes that are not valid UTF-8 (single cp1252 bytes such as 0x92): those bytes alone are decoded as cp1252 (latin-1 for the five bytes cp1252 leaves undefined) and meta.encoding says so; all other logs and all texts are plain UTF-8. Logs keep their CRLF line ends and every byte as released apart from the worker-ID replacement below. Worker IDs are replaced (owner decision 2026-10-13): the release keeps each writer's Amazon Mechanical Turk worker ID (21 values do not have the usual 'A' + 11-13 character form: 5 short numbers, longer codes, one e-mail address) in the Participant and __LogFileName fields of the writer's essay and typing-test IDFX logs, and some writers pasted or typed their ID into the essay (1 text) or during the tasks (it then appears in the CSV and IDFX logs). The writer's ID is read from the Participant field of their essay IDFX log (one distinct value per writer; the typing tests use '<ID>_<test>') and replaced by a code: 'W' + the writer's 4-digit rank in sorted release-ID order (10010800 -> W0001), padded with '_' to the length of the ID so that character positions in the logs stay valid, e.g. 'W0001_________' and 'W0001__________TYPING1'. The same code is used for that writer in every subset. IDs of 8 or more characters with a letter (all but 5) are replaced wherever they occur in the writer's essay text, holistic_scores_text, keystroke, typing-test and vocabulary-test logs, and so is any piece of 8 or more characters of such an ID (with a letter) found inside a run of capitals and digits (e.g. the part of a pasted ID that a writer then cut, which the logs record as removed text), with the code padded to the piece's length; the 5 short numeric Participant values (e.g. a 4-digit number) are replaced only in the two IDFX fields, as elsewhere they are ordinary numbers. No mapping is stored. meta.worker_id_replaced counts the replacements in a row. These are the only changes to texts and logs. A scan of the built files (case-insensitive) finds none of the original IDs and no piece of 8 or more of their characters. An ID typed key by key leaves one character per log event; such sequences are not reassembled. Not changed: an MTurk requester ID (the public ID of a requester account, not a worker) in a turkerview.com URL that one writer (83107247) pasted into the essay. Left out: WritingTask/writing_prompts.docx (its prompt texts are in `prompt`) and WritingTask/holistic_rating_form.docx (the raters' rubric), which are not records. Nothing is sampled or filtered.

*Licence:* MIT (LICENSE in the official GitHub repo, (c) 2025 terryyutian; the README links the Drive data, which state no other licence). The JoWR paper itself is CC BY-NC-ND 3.0. Privacy: the official release keeps the writers' Amazon Mechanical Turk worker IDs in the IDFX logs (and a few texts and CSV logs); here every one is replaced by an anonymous code (W0001...), see processing. MIT permits the modification.

### `l1_signals`

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).

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). `text` = `abstract`, `title` = `title`, both as released (verbatim, since some rows are CC BY-NC-ND); `date` = `year` (ACL files) or `publication_year` (arXiv file; its `year` is null); `source_record_id` = '<file>:<0-based line>' (the release has no id: DOIs repeat or are missing); `source_url` = `link` if given, else https://doi.org/<doi>; `meta` = 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 `quotation` (a short verbatim quote of a paper, non-commercial research use with credit). Nothing is left out.

*Licence:* 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.

### `longbench_write`

LongBench-Write (LongWriter). Bai, Zhang, Lv, Zheng, Zhu, Hou, Dong, Tang & Li (ICLR 2025).

`longbench_write` = every line of evaluation/longbench_write.jsonl, the 120-prompt LongBench-Write set that LongWriter's evaluation/pred.py reads (60 English, 60 Chinese; paper Sec. 3.3). The release's evaluation/longbench_write_en.jsonl holds the 60 English prompts again (identical lines), so they are not repeated: `meta.in_longbench_write_en` marks them. `longbench_write_ruler` = every line of evaluation/longwrite_ruler.jsonl, the LongWrite-Ruler probe of paper Sec. 2 (4 tasks in English and Chinese, each asking for 1,000 / 2,000 / 5,000 / 10,000 / 20,000 / 30,000 words or characters). The prompt is sent alone as the user turn (pred.py). `category` = the release's `type` (e.g. 'Literature and Creative Writing'); `meta` keeps `length` (the required output length) and the line index. No split in the release (evaluation-only), so the split is `full`. The prompts were collected by the authors (released 2024-08). LongWriter-6k (the separate SFT dataset with model-written answers) is not part of this benchmark and not included.

*Licence:* Apache-2.0 (repository LICENSE.txt; no separate data licence).

### `longbench_write_ruler`

LongBench-Write (LongWriter). Bai, Zhang, Lv, Zheng, Zhu, Hou, Dong, Tang & Li (ICLR 2025).

`longbench_write` = every line of evaluation/longbench_write.jsonl, the 120-prompt LongBench-Write set that LongWriter's evaluation/pred.py reads (60 English, 60 Chinese; paper Sec. 3.3). The release's evaluation/longbench_write_en.jsonl holds the 60 English prompts again (identical lines), so they are not repeated: `meta.in_longbench_write_en` marks them. `longbench_write_ruler` = every line of evaluation/longwrite_ruler.jsonl, the LongWrite-Ruler probe of paper Sec. 2 (4 tasks in English and Chinese, each asking for 1,000 / 2,000 / 5,000 / 10,000 / 20,000 / 30,000 words or characters). The prompt is sent alone as the user turn (pred.py). `category` = the release's `type` (e.g. 'Literature and Creative Writing'); `meta` keeps `length` (the required output length) and the line index. No split in the release (evaluation-only), so the split is `full`. The prompts were collected by the authors (released 2024-08). LongWriter-6k (the separate SFT dataset with model-written answers) is not part of this benchmark and not included.

*Licence:* Apache-2.0 (repository LICENSE.txt; no separate data licence).

### `masc`

MASC 3.0.0 and the Open American National Corpus (OANC). Ide, Baker, Fellbaum, Fillmore & Passonneau (ACL 2008); Ide, Fellbaum, Baker & Passonneau (ACL 2010); Ide & Suderman (LREC 2004, OANC).

Two releases of the ANC project, one subset each: `masc` = every document of MASC 3.0.0 (392: written 376 — blog, email incl. Enron, W3C and spam mail, essays, ficlets, fiction, govt-docs, jokes, journal, letters, movie-script, newspaper, non-fiction, technical, travel-guides, twitter (two files, each a stream of tweets) — and spoken 16: court-transcript, debate-transcript, face-to-face, telephone); `masc_oanc` = every document of the OANC GrAF release (8,824: written 6,424 — fiction, journal (Slate, Verbatim), letters (ICIC), non-fiction (OUP), technical (911report, biomed, government, plos), travel_guides (Berlitz) — and spoken 2,400: Charlotte face-to-face narratives and Switchboard telephone conversations). One row per primary text file `<doc>.txt`; split `full` (the releases have no splits). Nothing is sampled or filtered. MASC was drawn largely from OANC, so some documents occur in both subsets (the same text, or nearly; `text_hash` finds exact copies); both copies are kept, each under its own release. `text` = the .txt file as released (UTF-8; one leading byte-order mark removed). Many OANC files (Charlotte, ICIC letters, 911report, biomed, PLoS, Berlitz 1) indent every line and put blank lines between paragraphs or speaker turns; this layout is kept as released (the builder only removes trailing spaces and leading/trailing blank lines). Spoken documents are the release's transcripts as given. `source_record_id` = the archive member path without `.txt` (e.g. `MASC-3.0.0/data/written/blog/Acephalous-Cant-believe`). `genre` = '<modality>: <genre directory>' ('written: journal', 'spoken: telephone', ...; OANC's written_1/written_2 directories are both 'written'). `title` = the header's sourceDesc title, else its titleStmt title, skipping a title that only repeats the file name (headers sometimes give a corpus or publication name here, e.g. 'ICIC Corpus of Philanthropic Fundraising Discourse'; kept as given). `date` from the header's pubDate (its value attribute, else its text): an ISO date, 'Month D, YYYY', 'Month YYYY' or 'M/D/YY[YY] time' is read as a date (YYYY-MM-DD or YYYY-MM; a month or day of 00 is dropped), otherwise a single year in it gives YYYY; ranges ('2000-2002'), 'unknown' and missing pubDates (OANC government and Berlitz files) give no date. `meta` holds: corpus, archive, member (zip path of the .txt), header_member, modality, genre_dir, sub_dir, the header fields doc_id, header_title, source_title, authors, publisher, pub_name, pub_date_value, pub_date_text, cat_ref (the ANC text class, e.g. 'WR JO'), domain, subdomain, subject, medium, audience, language, extent_words, date_created, and header_xml (the whole header file as released). Left out: the stand-off annotation files (XML: sentences, tokens, POS, noun and verb chunks, named entities, Penn Treebank, FrameNet, WordNet senses, ...), MASC's original-annotations folder (third-party annotations in their own formats) and documentation; 250 OANC biomed headers (.anc) that have no primary text file in the release (nothing to include).

*Licence:* MASC: Creative Commons Attribution 3.0 United States (anc.org/data/masc: 'can be downloaded by anyone for any purpose, under the conditions of the Creative Commons Attribution 3.0 United States License'). OANC: 'unrestricted in terms of usage and redistribution' (anc.org/data/oanc). Both subsets are given CC BY 3.0 (the stricter of the two statements; corpus licence id cc-by-3.0 for the US port): credit the ANC project.

### `masc_oanc`

MASC 3.0.0 and the Open American National Corpus (OANC). Ide, Baker, Fellbaum, Fillmore & Passonneau (ACL 2008); Ide, Fellbaum, Baker & Passonneau (ACL 2010); Ide & Suderman (LREC 2004, OANC).

Two releases of the ANC project, one subset each: `masc` = every document of MASC 3.0.0 (392: written 376 — blog, email incl. Enron, W3C and spam mail, essays, ficlets, fiction, govt-docs, jokes, journal, letters, movie-script, newspaper, non-fiction, technical, travel-guides, twitter (two files, each a stream of tweets) — and spoken 16: court-transcript, debate-transcript, face-to-face, telephone); `masc_oanc` = every document of the OANC GrAF release (8,824: written 6,424 — fiction, journal (Slate, Verbatim), letters (ICIC), non-fiction (OUP), technical (911report, biomed, government, plos), travel_guides (Berlitz) — and spoken 2,400: Charlotte face-to-face narratives and Switchboard telephone conversations). One row per primary text file `<doc>.txt`; split `full` (the releases have no splits). Nothing is sampled or filtered. MASC was drawn largely from OANC, so some documents occur in both subsets (the same text, or nearly; `text_hash` finds exact copies); both copies are kept, each under its own release. `text` = the .txt file as released (UTF-8; one leading byte-order mark removed). Many OANC files (Charlotte, ICIC letters, 911report, biomed, PLoS, Berlitz 1) indent every line and put blank lines between paragraphs or speaker turns; this layout is kept as released (the builder only removes trailing spaces and leading/trailing blank lines). Spoken documents are the release's transcripts as given. `source_record_id` = the archive member path without `.txt` (e.g. `MASC-3.0.0/data/written/blog/Acephalous-Cant-believe`). `genre` = '<modality>: <genre directory>' ('written: journal', 'spoken: telephone', ...; OANC's written_1/written_2 directories are both 'written'). `title` = the header's sourceDesc title, else its titleStmt title, skipping a title that only repeats the file name (headers sometimes give a corpus or publication name here, e.g. 'ICIC Corpus of Philanthropic Fundraising Discourse'; kept as given). `date` from the header's pubDate (its value attribute, else its text): an ISO date, 'Month D, YYYY', 'Month YYYY' or 'M/D/YY[YY] time' is read as a date (YYYY-MM-DD or YYYY-MM; a month or day of 00 is dropped), otherwise a single year in it gives YYYY; ranges ('2000-2002'), 'unknown' and missing pubDates (OANC government and Berlitz files) give no date. `meta` holds: corpus, archive, member (zip path of the .txt), header_member, modality, genre_dir, sub_dir, the header fields doc_id, header_title, source_title, authors, publisher, pub_name, pub_date_value, pub_date_text, cat_ref (the ANC text class, e.g. 'WR JO'), domain, subdomain, subject, medium, audience, language, extent_words, date_created, and header_xml (the whole header file as released). Left out: the stand-off annotation files (XML: sentences, tokens, POS, noun and verb chunks, named entities, Penn Treebank, FrameNet, WordNet senses, ...), MASC's original-annotations folder (third-party annotations in their own formats) and documentation; 250 OANC biomed headers (.anc) that have no primary text file in the release (nothing to include).

*Licence:* MASC: Creative Commons Attribution 3.0 United States (anc.org/data/masc: 'can be downloaded by anyone for any purpose, under the conditions of the Creative Commons Attribution 3.0 United States License'). OANC: 'unrestricted in terms of usage and redistribution' (anc.org/data/oanc). Both subsets are given CC BY 3.0 (the stricter of the two statements; corpus licence id cc-by-3.0 for the US port): credit the ANC project.

### `medquad`

MedQuAD: Medical Question Answering Dataset. Ben Abacha & Demner-Fushman, A Question-Entailment Approach to Question Answering (BMC Bioinformatics 2019).

Every question-answer pair of the 11,274 XML files of the 12 sub-collections (47,457 pairs). One row per <QAPair>: `text` = the <Answer> text (a section of an NIH/CDC consumer-health web page, written by the agencies' staff, extracted by the authors' patterns), `prompt` = the <Question> (the authors generated the questions from templates per question type and focus, e.g. 'What are the symptoms of X ?'; they are not consumer questions). The authors removed the answers of the three MedlinePlus subsets (10_MPlus_ADAM_QA 17,348, 11_MPlusDrugs_QA 12,889, 12_MPlusHerbsSupplements_QA 792 pairs) for copyright, and 5 GARD pairs have an empty answer in the release: these 31,034 pairs have empty text and are counted by the builder as dropped_empty (16,423 rows remain). `source_record_id` = '<sub-collection>/<file stem>/<pair number>' (pair number = the QAPair pid; position in the file if a pid is missing or repeated); `parent_id` = '<sub-collection>/<file stem>' (the web page; pairs of the same page share it); `source_url` = the page URL; `title` = None; `genre` = 'consumer health information (NIH web page section)'. `meta`: sub_collection, site, file, document_id, source, url, focus, the focus annotations (category, CUIs, semantic types, semantic group, synonyms), pid, qid, qtype, root_tag (three XML schemas occur: <Document> in 11,269 files, <DiseaseFile> in 9_CDC_QA/0000397.xml, lower-case <doc>/<pair> in 4 NINDS files). No per-page dates are released: `date` = '2019-01', the month the release was published (first commit 2019-01-25; pages crawled before), an upper bound. Left out: the LiveQA judged-answer file (QA-TestSet-LiveQA-Med-Qrels-2479-Answers.zip: relevance judgments, no text), readme and licence.

*Licence:* CC BY 4.0 (readme.txt: 'The MedQuAD dataset is published under a Creative Commons Attribution 4.0 International Licence'; LICENSE.txt). The answers are U.S. government (NIH/CDC) web text.

### `methods`

anagram-bench generation methods. anagram-bench (2026); each row cites the paper or repository it quotes.

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 `template`, or `classifier` for judge/filter prompts such as pv_rw0305_compliance_judge or mine_classifier_v1); every item of an instruction pool is a row (kind `instruction`, `name` = pool id, `meta.item_index` = position in the pool); every attack is a row (kind `attack`, `text` null, `params` = its settings as JSON, `meta` = code_url, revision, entry_point). A template used by two tasks appears once per task (`task`). `text` and `system` 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 `meta.text_turns`. `verbatim` is the recipe's flag (instruction items: true unless the pool is our own or a paraphrase of API documentation). `paper`/`paper_url` come from the build's work list (works_813.json); `location` is the recipe's provenance location (attacks: the entry point). Template settings (params, generation_params) go to `params`; all other recipe fields go to `meta` unchanged; pool-level fields go to every item's `meta`, 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; `meta.licence_basis` 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 `quotation` (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), `quotation` 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 `facts`. 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 `meta.embedded_source` 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 `meta.removed_passages`, sha256 of the original text in `meta.original_text_sha256`). PAID SERVICES (StealthGPT, Undetectable AI, Humbot, BypassGPT, QuillBot, DAMAGE's web tools) are given by name and settings only, `meta.note` = 'paid service; not run': the vendors' descriptions of their modes are cut from the mode lists (rw0389_quillbot_paraphraser_modes, rw0278_vendor_modes: `text` = the mode name, `meta.vendor_description_removed`) and from the Humbot/BypassGPT attack settings (`meta.params_trimmed`), 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.

*Licence:* Per row, in `license`: the licence of the quoted paper, repository or dataset, `quotation` for text whose source carries no open licence (non-commercial research use with attribution), `facts` for settings, and CC BY 4.0 for anagram-bench's own wording. Cite the paper in `paper` / `paper_url` when using a row.

### `million_authors`

The Million Authors Corpus (MAC). Meisenbacher, Nissen, Jurgens et al. (Findings of ACL 2025).

million_authors: every text chunk of every revision in every one of the 1,965 .jsonl.bz2 files of the full release (60 language folders, one Wikipedia edition each, dumps of 2023-06-01 for English and 2024 for the others), one row per chunk, split 'full': 60,083,121 rows (the paper's 60.08M chunks by 1.29M authors). Each line of a file is one revision; its `new_text` list holds the chunks of contiguous text the edit added (MAC's extraction: wiki markup mostly stripped, one sentence or paragraph per line); `text` = one chunk as released. `source_record_id` = '<wiki>:<revision id>:<chunk index>'; `author_id` = '<wiki>:<user id>' (the editor's local account id on that wiki; user name in meta); `title` = page title; `date` = revision timestamp; `genre` from the page namespace (Wikipedia article, talk page, user page, ...); `source_url` = the revision's permanent link. `meta` = wiki (language edition code, from the file name), dump (e.g. enwiki-20230601), folder (the release's language folder; note that folder 'bn' holds the Bashkir edition, bawiki-20240401, not Bengali: these rows have wiki 'ba'), file (member of the zip), revision fields as released (rev_id, parent_id, timestamp, user {id, text}, page {id, title, namespace, restrictions}, minor, comment, bytes, sha1, deleted, detected_langs) and chunk_index / n_chunks. Revisions with no chunk add no row. Left out of the full release: the 60 metadata_<lang>.csv files (per-page extraction statistics: page id, namespace, dump file, revisions read and saved, bytes, runtime; no text). million_authors_av: every text of the concise release, for every language and split: <language>/<lang>_wikipedia/{train,dev,test}_{queries,candidates}.jsonl.gz, one row per line, split = train / dev / test (the authors' 7:1:2 split by editor; 1,889,960 / 270,020 / 540,054 rows), meta.role = query or candidate. These texts are chunks of the full release, re-normalised by the authors (whitespace, list markers, foreign leading lines trimmed), so most duplicate rows of million_authors (match on meta.sourceSpecific user_id + page_id + dateCreated/timeCreated). `source_record_id` = documentID; `author_id` = the release's author UUID (authorIDs); `date` = dateCreated + timeCreated; `title` = page title; `meta` = every other field as released (authorIDs, spanAttribution, collectionNum, source, dateCollected, publiclyAvailable, deidentified, languages, lengthWords, dateCreated, timeCreated, isForeground, sourceSpecific) plus language folder and file. The `source` field says 'ru.wikipedia.org' for every language (a release bug; kept). The per-language corpus.jsonl.gz files are empty. Nothing else is left out; no filtering of bots, templates or boilerplate.

*Licence:* Zenodo record: CC BY 4.0; Hugging Face card: CC BY-SA 4.0; the texts are Wikipedia contributions (CC BY-SA). The stricter CC BY-SA 4.0 is used for every row: attribution (the editors, through the user and revision ids given per row, and the MAC authors) and share-alike.

### `million_authors_av`

The Million Authors Corpus (MAC). Meisenbacher, Nissen, Jurgens et al. (Findings of ACL 2025).

million_authors: every text chunk of every revision in every one of the 1,965 .jsonl.bz2 files of the full release (60 language folders, one Wikipedia edition each, dumps of 2023-06-01 for English and 2024 for the others), one row per chunk, split 'full': 60,083,121 rows (the paper's 60.08M chunks by 1.29M authors). Each line of a file is one revision; its `new_text` list holds the chunks of contiguous text the edit added (MAC's extraction: wiki markup mostly stripped, one sentence or paragraph per line); `text` = one chunk as released. `source_record_id` = '<wiki>:<revision id>:<chunk index>'; `author_id` = '<wiki>:<user id>' (the editor's local account id on that wiki; user name in meta); `title` = page title; `date` = revision timestamp; `genre` from the page namespace (Wikipedia article, talk page, user page, ...); `source_url` = the revision's permanent link. `meta` = wiki (language edition code, from the file name), dump (e.g. enwiki-20230601), folder (the release's language folder; note that folder 'bn' holds the Bashkir edition, bawiki-20240401, not Bengali: these rows have wiki 'ba'), file (member of the zip), revision fields as released (rev_id, parent_id, timestamp, user {id, text}, page {id, title, namespace, restrictions}, minor, comment, bytes, sha1, deleted, detected_langs) and chunk_index / n_chunks. Revisions with no chunk add no row. Left out of the full release: the 60 metadata_<lang>.csv files (per-page extraction statistics: page id, namespace, dump file, revisions read and saved, bytes, runtime; no text). million_authors_av: every text of the concise release, for every language and split: <language>/<lang>_wikipedia/{train,dev,test}_{queries,candidates}.jsonl.gz, one row per line, split = train / dev / test (the authors' 7:1:2 split by editor; 1,889,960 / 270,020 / 540,054 rows), meta.role = query or candidate. These texts are chunks of the full release, re-normalised by the authors (whitespace, list markers, foreign leading lines trimmed), so most duplicate rows of million_authors (match on meta.sourceSpecific user_id + page_id + dateCreated/timeCreated). `source_record_id` = documentID; `author_id` = the release's author UUID (authorIDs); `date` = dateCreated + timeCreated; `title` = page title; `meta` = every other field as released (authorIDs, spanAttribution, collectionNum, source, dateCollected, publiclyAvailable, deidentified, languages, lengthWords, dateCreated, timeCreated, isForeground, sourceSpecific) plus language folder and file. The `source` field says 'ru.wikipedia.org' for every language (a release bug; kept). The per-language corpus.jsonl.gz files are empty. Nothing else is left out; no filtering of bots, templates or boilerplate.

*Licence:* Zenodo record: CC BY 4.0; Hugging Face card: CC BY-SA 4.0; the texts are Wikipedia contributions (CC BY-SA). The stricter CC BY-SA 4.0 is used for every row: attribution (the editors, through the user and revision ids given per row, and the MAC authors) and share-alike.

### `mt_bench`

Chatbot Arena Conversations (MT-Bench and Chatbot Arena paper). Zheng, Chiang, Sheng, Zhuang, Wu, Zhuang, Lin, Li, Li, Xing, Zhang, Gonzalez & Stoica (NeurIPS 2023 Datasets and Benchmarks).

The release of the MT-Bench / Chatbot Arena paper: 33,000 conversations that users of Chatbot Arena (chat.lmsys.org) had with two anonymous models side by side between 2023-04-24 and 2023-06-26, with the user's vote; one split, `train`. Every conversation is kept (all languages, flagged ones included). Both sides of a battle share the user turns (checked on all rows), so `messages` holds conversation A's turns with each assistant turn's `content` set to null: the model outputs are CC BY-NC 4.0, while these rows are CC BY 4.0 (user prompts only); conversation B adds nothing else and is left out. `model` is null because two models answered; `meta` keeps the release's other fields as released (`question_id`, `model_a`, `model_b`, `winner`, `judge`, `turn`, `anony`, `language`, `tstamp`, `openai_moderation`, `toxic_chat_tag`). `date` is `tstamp` in UTC. `question_id` occurs more than once in 19 cases; later copies get `#2`, `#3` in release order. The 80 MT-Bench questions (written by the authors, not users) and the MT-Bench judgment files are separate releases and not part of this subset.

*Licence:* README: "The user prompts are licensed under CC-BY-4.0, while the model outputs are licensed under CC-BY-NC-4.0" (card tag `cc`). Gated with automatic approval; the README's Disclaimers and Terms ask users not to try to identify individuals, to use the data lawfully, and to follow the model providers' terms for model outputs (not included here). The data may contain unsafe or offensive text.

### `muldiloc`

MulDiLoC-HAP: Multidialectal, Longitudinal Corpus of Human-AI Hybrid Production (human-written essays). Gan, Adams, Dunn & Nini (LREC 2026).

MulDiLoC-HAP.csv has 862 rows: 682 consenting participants (Condition Nat 147, Sty 179, Phr 179, Par 177; varieties US, UK, CAN, AUS, NZ) and 180 GPT-4.1 persona rows (Condition GPT), each with up to three essays (Essay1-3, Topic1-3). Session 1 is natural writing without assistance; in session 2 condition Nat writes without assistance while Sty (grammar and style feedback), Phr (incremental suggestions) and Par (continuing an AI-written opening) write with GPT-4.1; session 3 is an unassisted delayed posttest three weeks later. One row per human-written essay, 1,417 in all: every session-1 essay of the 682 participants (682), the session-2 essays of condition Nat (147) and every session-3 essay (588; 94 participants have none). Left out: the 535 session-2 essays of conditions Sty (179), Phr (179) and Par (177), written with LLM assistance (human-AI hybrid text), and the 540 essays of the 180 GPT rows (AI-generated). The release's other files (measures.csv, lexeme.csv, nominalisation_pos.csv: per-essay measures and aggregate counts; scripts) are not texts and are not used. `source_record_id` = '<ID>:s<session>'; `author_id` = the participant ID (links a writer's essays); `text` = Essay<session> as released (curly quotes that appear as mojibake such as '’' in some essays are kept); `prompt` = the topic text shown to writers for the essay's topic label (Topic<session>: stereotype, disagreement, technology, dream, workplace), from the study interface's topic_prompts.js (keys a-e, matched to the labels by content); `genre` 'informal forum-style personal essay'; `date` = 2025 (the study ran in 2025: topics added to the interface 2025-05, CSV committed 2025-10). `meta`: session, session_task, Topic<session> label, Condition, Variety and every participant field of the CSV as released (age, gender, education_org, education, occupation, ethnicity_org, ethnicity, englishVariety, otherLanguages, currentLocation, previousLocations, aiFamiliarity, aiTools, Otherai, aiFrequency, aiPurposes, Otheraipurpose, aiOpinion, consent).

*Licence:* MIT (LICENSE file of the release, (c) 2025 Qiao Gan, Benjamin Adams, Jonathan Dunn, Andrea Nini). README: the release includes only the 682 participants who consented to share their work publicly. The topic texts come from the MIT-licensed study interface (coauthor-interface, (c) 2023 Mina Lee).

### `multieurlex`

MultiEURLEX: multilingual EU legislation with EUROVOC labels. Chalkidis, Fergadiotis & Androutsopoulos (EMNLP 2021).

Every law of all three splits (train 55,000 laws 1958-2010, validation 5,000 2010-2012, test 5,000 2012-2016; the release's dev.jsonl is the card's 'validation' split) in every one of the 23 languages the release holds: one row per law per language, 1,044,739 rows (train 817,239, validation 112,500, test 115,000; a law missing in a language has no row there: laws adopted before a country joined the EU often lack its language). `text` = the language version as released (the authors' plain text of the EUR-Lex HTML: title, citations, recitals, articles, annexes; line breaks as released). `source_record_id` = '<celex_id>/<language code>'; `parent_id` and meta.parallel_id = the CELEX number, the same for all language versions of a law; `date` = the release's publication_date (the date of the act); `source_url` = the act's EUR-Lex page in that language; `genre` 'EU legislation'. `meta`: parallel_id, release_language (the release's language key; the `language` column is the builder's language id of the text), eurovoc_concepts as released (level_1, level_2, level_3, all_levels: EUROVOC concept ids). Rows are in file order (law by law, languages in the order above). Nothing is left out.

*Licence:* Stated inconsistently: the HF card's YAML says cc-by-sa-4.0, while the card's licensing text and the Zenodo record say CC BY 4.0 ('We provide MultiEURLEX with the same licensing as the original EU data (CC-BY-4.0): (c) European Union, 1998-2021'; EUR-Lex reuse under Commission Decision 2011/833/EU). The stricter reading, CC BY-SA 4.0, is used for every row: attribution to the authors and to the European Union, and share-alike.

### `newswire`

Newswire: A Large-Scale Structured Database of a Century of Historical News. Silcock, Arora, D'Amico-Wong & Dell (NeurIPS 2024 Datasets and Benchmarks).

Every article of all 100 year files (1878-1977), in file order; nothing sampled or filtered. The release is already one record per unique wire article (the authors kept one copy per cluster of newspaper reprints, with the reprinting newspapers listed). `text` = `article` (OCR text of the wire article as released, OCR errors and line breaks kept). `source_record_id` = '<year>_<position in the year file, from 0>' (the release has no article ids). `date` = the earliest of `dates` (the publication dates of the reprints, mmm-DD-YYYY) as YYYY-MM-DD, else `year`; `title`, `author_id`, `parent_id` and `source_url` are empty (the release has no headlines, author ids or per-article URLs; `byline` is in meta). `genre` = 'newswire article (historical U.S. newspapers, OCR)'. `meta` holds every other field as released: year, dates, byline, newspaper_metadata (lccn, title, city, state of each newspaper that ran the article), the seven topic flags antitrust, civil_rights, crime, govt_regulation, labor_movement, politics, protests, ca_topic (Comparative Agendas topic), ner_words and ner_labels (token-level NER), wire_city, wire_state, wire_country, wire_coordinates, wire_location_notes, people_mentioned (wikidata_id, person_name, person_gender, person_occupation) and cluster_size; plus file. Left out: nothing of the pinned release; all 2,719,607 articles (the count the paper gives) have text and are rows. The `cleaned_article` field that the later repo head adds is a GPT-4o-mini rewrite (AI-edited) and is not used. Every article has at least one parseable date (1878-01-01 .. 1977-12-31).

*Licence:* CC BY 4.0 (dataset card 'license: cc-by-4.0'; README: 'The dataset has a CC-BY 4.0 license'). Attribution required. The underlying wire articles (1878-1977) are described by the authors as public domain; the paper's datasheet links the CC BY 2.0 deed, also an attribution-only licence. Historical news may contain period offensive language and details of people named in the news.

### `nirvana`

NIRVANA: A Comprehensive Dataset for Reproducing How Students Use Generative AI for Essay Writing. Jelson, Manesh, Lee, Jang, Dunlap, Maddox, Kim & Lee (arXiv 2026).

The release is one workbook, dataset.xlsx, with two sheets: 'Participants' (77 rows, one per student/essay: demographics, final essay text, editing statistics and survey scores) and 'Essay Data' (26,142 logged events). Each session (essay_num) becomes one conversation: its `gpt_inquiry` events (the student's message to ChatGPT) are the user turns and its `gpt_response` events (ChatGPT's reply) the assistant turns, in op_index order (every session alternates inquiry/response). 71 of the 77 sessions used ChatGPT; the 6 sessions without any inquiry have no conversation and give no row. In essay 69 the release repeats a block of 30 events (same op_index, time and text), which contains one inquiry/response pair; the copy is not repeated in `messages`, so the 71 conversations hold 347 inquiries and 347 replies (348 of each in the log). `prompt` = the first inquiry; `model` = gpt-3.5-turbo (paper Sec. 3.1). `meta.turns` keeps, per message, the event's xlsx index, op_index, op_type and `time` (seconds since session start; the release has no absolute dates, so `date` is null); `meta.participant` keeps that student's 'Participants' row as released (including 'Final Text', the essay they submitted, which mixes their own and ChatGPT's text). Not included: the rest of the 'Essay Data' event log (editor keystroke/insert/delete/paste/cursor events and copy/paste events on the ChatGPT panel, 25,446 events). They are edit operations, not prompts or conversation turns, so they fit none of the record types; they remain in the official workbook.

*Licence:* CC BY 4.0 (OSF project licence 'CC-By Attribution 4.0 International'); it covers the whole release, including the logged ChatGPT replies.

### `parlamint`

ParlaMint 5.0: comparable corpora of parliamentary debates (29 parliaments). Erjavec, Kopp, Ogrodniczuk, Osenova et al. (Language Resources and Evaluation 2023; ParlaMint 5.0, CLARIN.SI 2025).

Every speech of the 29 corpora of ParlaMint 5.0 (CLARIN.SI hdl 11356/2004), one row per speech (TEI utterance <u>): 8,073,406 speeches in 46,204 sitting files, exactly the per-corpus counts of the release's speech statistics (Build/Metadata/ParlaMint-speeches-stats.tsv in ParlaMint-5.0.tgz) (AT 231,759, BA 126,252, BE 199,305, BG 210,018, CZ 196,185, DK 398,610, EE 227,872, ES 76,369, ES-CT 50,824, ES-GA 83,078, ES-PV 39,148, FI 146,858, FR 714,860, GB 670,912, GR 342,274, HR 504,338, HU 116,346, IS 95,286, IT 172,796, LV 162,782, NL 609,248, NO 398,809, PL 228,326, PT 248,577, RS 316,069, SE 84,662, SI 311,354, TR 681,052, UA 429,437). Sittings of 1996-01-15 to 2024-03-13 (most corpora 2014/15 to 2022; AT from 1996, RS 1997, NO and BA 1998, SI 2000, UA 2002, HR 2003; the 93,951 speeches of CZ, ES, HU, PT and UA dated 2022-11-30 or later have pre_chatgpt = false). Languages: each corpus's own (bs, bg, ca, cs, da, de, el, en, es, et, eu, fi, fr, gl, hr, hu, is, it, lv, nl, nb, nn, pl, pt, ru, sl, sr, sv, tr, uk); multilingual parliaments (BE, ES-CT, ES-PV, FI, NO, UA) keep each speech in the language it was given in. Nothing is sampled or filtered and no speech is empty. `text` = the release's own plain text of the speech (ParlaMint-XX.txt/, made by Scripts/parlamint-tei2text.xsl: the speech's text, transcriber comments inside it (<note>, <gap>, <vocal>, <kinesic>, <incident>: applause, interruptions, time stamps, omitted material) written as [[...]], whitespace normalised), except that the boundaries between its paragraphs (<seg>) and between-paragraph comments, which the plain text joins with a space, are newlines: the text is rebuilt from the TEI with the same rules, and for every speech it equals the official plain-text line once newlines are replaced by spaces (checked for all 8,073,406). The texts are the parliaments' official transcripts as ParlaMint encoded them. `source_record_id` = the speech's xml:id; `parent_id` = the sitting id (Text_ID); `date` = the sitting date; `author_id` = Speaker_ID (ParlaMint person id; empty when '-'); `title` = the <head>(s) of the enclosing debate section joined by ' | ', where the corpus has them (DK, ES, ES-CT, FI, FR, GB, GR, IT, NL, NO, PL, PT, SE, SI), else empty; `source_url` = the sitting's source document at the parliament (TEI idno subtype 'parliament'). meta = corpus, the speech's row of <sitting>-meta-en.tsv as released (Title = the sitting title, Date, Body, Term, Session, Meeting, Sitting, Agenda, Subcorpus (Reference, COVID or 'COVID,War'), Lang, Speaker_role, Speaker_MP, Speaker_minister, Speaker_party, Speaker_party_name, Party_status, Party_orientation, Speaker_ID, Speaker_name, Speaker_gender, Speaker_birth, Topic (CAP topic); '-' = not given), `local` = the values of <sitting>-meta.tsv (corpus-language names) that differ from the English ones, section_heads, and `u` = the utterance's other TEI attributes (who = speaker reference, missing for 66,132 unidentified speakers; ana = speaker type and topic; n in BA, BG, HR, RS; xml:lang in BE, ES-CT, NO; source in CZ, IS). No translations are in this release, so there is no parallel_id. Left out: the separately distributed linguistically annotated edition (ParlaMint.ana 5.0, hdl 11356/2005) and English edition machine-translated from all corpora (ParlaMint-en.ana 5.0, hdl 11356/2006); transcriber comments that stand between speeches (outside any <u>: procedural notes, time stamps, incidents), which belong to no speech; the TEI corpus root files (corpus header, person and organisation lists, taxonomies), whose information reaches the rows through the TSV metadata; ParlaMint-5.0.tgz (a snapshot of the project's GitHub repository: scripts, schemas, samples that repeat corpus text, build metadata) and ParlaMint-logs.tgz (build logs).

*Licence:* CC BY 4.0 (CLARIN.SI item licence and the <licence> of every corpus's TEI header). The GB sittings also state the Open Parliament Licence v3.0 ("Contains Parliamentary information licensed under the Open Parliament Licence v3.0"). Cite the ParlaMint 5.0 handle and Erjavec et al. (2023).

### `parliament_undisclosed_llm`

Detecting undisclosed LLM-generated content in parliamentary texts (data release). Suvanto, McGlinchey, Barclay & Wahde (arXiv 2026).

The release (one zip, 17,780 text files) holds UK and Swedish parliamentary texts. Base sets (UK/Base and SWE/Base, files Training/Validation/Test -> splits train/validation/test): one paragraph of >= 30 tokens per line as '<label>\t<text>', shuffled; label 0 = the human original (UK written ministerial statements of 2014-2020 from Hansard; Swedish Riksdag members' motions of sessions 2014/15-2020/21), label 1 = a paragraph of an LLM-generated text (the authors summarised 1,400 human texts per language with Gemini 3.1 Flash-Lite and regenerated full texts from the summaries with Gemini 3.1 Flash-Lite and, for the UK, GPT-5 mini). Only the label-0 lines are included, every one of them: UK train 7,653, validation 957, test 957; Swedish train 6,274, validation 784, test 785 (17,410 rows). Left out: the 13,342 label-1 lines (UK 6,772, Swedish 6,570), which are LLM text; and the whole unlabelled forward test set (UK/Forward/<year>/*.txt, 4,209 UK written statements of 2021 to 2026-04-29; SWE/Forward/<session>/*.txt, 13,565 Riksdag motions of sessions 2021/22 to 2025/26), because its authorship is uncertain: the texts carry no label and the paper reports undisclosed LLM use in them rising from 2022 on. Text is as released, which is NOT the original wording: the authors tokenised every text (split on punctuation and whitespace, punctuation marks removed, URLs replaced by 'URLPLACEHOLDER', case kept), so it has no punctuation; CRLF line ends are removed. The original wording is in Hansard (hansard.parliament.uk) and the Riksdag open data, not in this release. Swedish texts had headers and signatures removed by the authors and were length-filtered (2,000-5,000 characters) before being cut into paragraphs. Lines carry no statement / motion ids, so paragraphs of one text cannot be regrouped. `source_record_id` = '<file path>:<line number>' (1-based); `language` comes from the builder (en / sv); `genre` 'written ministerial statement' (UK) or 'parliamentary motion' (Swedish). `date` is an upper bound, since the release gives no dates: UK 2020 (statements of 2014-2020), Swedish 2021 (sessions 2014/15-2020/21). `meta`: parliament (UK / SWE), set (Base), file, line, label (0), period.

*Licence:* CC BY 4.0 (Zenodo record). UK texts: Hansard, 'Contains Parliamentary information licensed under the Open Parliament Licence v3.0'. Swedish texts: Riksdag open data under the Riksdag's terms of use (public documents). Cite Suvanto et al. (2026) and credit both parliaments.

### `peerread`

PeerRead: A Dataset of Peer Reviews (ACL 2017 and CoNLL 2016 sections). Kang, Ammar, Dalvi, van Zuylen, Kohlmeier, Hovy & Schwartz (NAACL 2018).

PeerRead v1.0 has five kinds of section: acl_2017 and conll_2016 (submissions whose author opted in the draft and at least one reviewer opted in an anonymous review, via softconf), iclr_2017 (OpenReview reviews), nips_2013-2017 (not in the release: crawl script only) and three arXiv sections (arxiv.cs.ai/cs.cl/cs.lg_2007-2017: arXiv papers with accept labels, no review text). Only acl_2017 and conll_2016 are used, because only they carry an explicit open licence (data/acl_2017/LICENSE.txt and data/conll_2016/LICENSE.txt, the CC BY 4.0 legal code, added in commit 954a920 'add license info'; the paper, Sec. 2.3: 'different sections may have different license agreements'). Left out: iclr_2017 (427 papers, OpenReview reviews and drafts; no licence file, owner decision), the arXiv sections (11,778 papers; no licence file, and the per-paper arXiv licences are not recorded in the release; they hold no review text: every 'reviews' list is empty), nips_2013-2017 (not released), data/acl_accepted.txt (a list of 19,743 ACL Anthology titles used by the authors' code to label acceptance; outside the licensed sections), code/ and its aspect annotation file, and the PDFs (pdfs/<id>.pdf; their text is the release's science-parse JSON, which is used). All three official splits of both sections (train/dev/test, as released; ACL 2017: 123/7/7 papers, CoNLL 2016: 19/1/2). peerread: one row per entry of a paper's 'reviews' list (reviews/<id>.json), in list order (papers in numeric id order; 314 reviews = the paper's 275 ACL + 39 CoNLL); all entries are reviewer-written softconf review forms (is_meta_review is null for all; no entry repeats within a list). `text` = `comments` (hard-wrapped at about 80 characters as exported from softconf; a few mojibake sequences as released). `meta` = every other field of the review as released (strings: RECOMMENDATION, REVIEWER_CONFIDENCE, IMPACT, SUBSTANCE, APPROPRIATENESS, MEANINGFUL_COMPARISON, SOUNDNESS_CORRECTNESS, ORIGINALITY, CLARITY, REPLICABILITY (CoNLL), PRESENTATION_FORMAT, is_meta_review, ...) plus venue, split, paper_id, paper_title, review_index. `source_record_id` = '<venue>/<split>/<paper id>/r<index>'; `parent_id` = the id of the paper's row in peerread_papers. peerread_papers: one row per paper (reviews/<id>.json with parsed_pdfs/<id>.pdf.json). `text` = the science-parse text of the draft as released: metadata.abstractText, then each entry of metadata.sections as its heading (when not null) on its own line followed by its text, blocks separated by a blank line; nothing is cleaned (the drafts are anonymous review copies with softconf line numbers, and 157 of the 159 texts keep runs of them such as '000', '011', '012' on their own lines, as science-parse extracted them). 159 papers (the paper's 137 ACL + 22 CoNLL), all with a parsed PDF. `title` = the paper JSON's title. `meta` = the paper JSON's other fields as released (abstract, id, and histories (ACL, always empty); the release gives no authors or accept decisions for these sections), n_reviews, and the science-parse record minus the text used above (name, metadata.source, title, authors, emails, references, referenceMentions, year, creator) under meta.science_parse. `date` = the venue year (2017 / 2016: submission and review period; the release has no per-item dates). No author ids (reviews and drafts are anonymous). Text is human-written; the release has no AI-generated text.

*Licence:* CC BY 4.0 (data/acl_2017/LICENSE.txt and data/conll_2016/LICENSE.txt in the release: the Creative Commons Attribution 4.0 International legal code, covering each section's reviews and paper drafts, which the reviewers and authors opted in for this dataset). Credit PeerRead (Kang et al., NAACL 2018) and the ACL 2017 / CoNLL 2016 reviewers and authors. The repository itself has no top-level licence; the iclr_2017 and arXiv sections carry none and are not included.

### `peerread_papers`

PeerRead: A Dataset of Peer Reviews (ACL 2017 and CoNLL 2016 sections). Kang, Ammar, Dalvi, van Zuylen, Kohlmeier, Hovy & Schwartz (NAACL 2018).

PeerRead v1.0 has five kinds of section: acl_2017 and conll_2016 (submissions whose author opted in the draft and at least one reviewer opted in an anonymous review, via softconf), iclr_2017 (OpenReview reviews), nips_2013-2017 (not in the release: crawl script only) and three arXiv sections (arxiv.cs.ai/cs.cl/cs.lg_2007-2017: arXiv papers with accept labels, no review text). Only acl_2017 and conll_2016 are used, because only they carry an explicit open licence (data/acl_2017/LICENSE.txt and data/conll_2016/LICENSE.txt, the CC BY 4.0 legal code, added in commit 954a920 'add license info'; the paper, Sec. 2.3: 'different sections may have different license agreements'). Left out: iclr_2017 (427 papers, OpenReview reviews and drafts; no licence file, owner decision), the arXiv sections (11,778 papers; no licence file, and the per-paper arXiv licences are not recorded in the release; they hold no review text: every 'reviews' list is empty), nips_2013-2017 (not released), data/acl_accepted.txt (a list of 19,743 ACL Anthology titles used by the authors' code to label acceptance; outside the licensed sections), code/ and its aspect annotation file, and the PDFs (pdfs/<id>.pdf; their text is the release's science-parse JSON, which is used). All three official splits of both sections (train/dev/test, as released; ACL 2017: 123/7/7 papers, CoNLL 2016: 19/1/2). peerread: one row per entry of a paper's 'reviews' list (reviews/<id>.json), in list order (papers in numeric id order; 314 reviews = the paper's 275 ACL + 39 CoNLL); all entries are reviewer-written softconf review forms (is_meta_review is null for all; no entry repeats within a list). `text` = `comments` (hard-wrapped at about 80 characters as exported from softconf; a few mojibake sequences as released). `meta` = every other field of the review as released (strings: RECOMMENDATION, REVIEWER_CONFIDENCE, IMPACT, SUBSTANCE, APPROPRIATENESS, MEANINGFUL_COMPARISON, SOUNDNESS_CORRECTNESS, ORIGINALITY, CLARITY, REPLICABILITY (CoNLL), PRESENTATION_FORMAT, is_meta_review, ...) plus venue, split, paper_id, paper_title, review_index. `source_record_id` = '<venue>/<split>/<paper id>/r<index>'; `parent_id` = the id of the paper's row in peerread_papers. peerread_papers: one row per paper (reviews/<id>.json with parsed_pdfs/<id>.pdf.json). `text` = the science-parse text of the draft as released: metadata.abstractText, then each entry of metadata.sections as its heading (when not null) on its own line followed by its text, blocks separated by a blank line; nothing is cleaned (the drafts are anonymous review copies with softconf line numbers, and 157 of the 159 texts keep runs of them such as '000', '011', '012' on their own lines, as science-parse extracted them). 159 papers (the paper's 137 ACL + 22 CoNLL), all with a parsed PDF. `title` = the paper JSON's title. `meta` = the paper JSON's other fields as released (abstract, id, and histories (ACL, always empty); the release gives no authors or accept decisions for these sections), n_reviews, and the science-parse record minus the text used above (name, metadata.source, title, authors, emails, references, referenceMentions, year, creator) under meta.science_parse. `date` = the venue year (2017 / 2016: submission and review period; the release has no per-item dates). No author ids (reviews and drafts are anonymous). Text is human-written; the release has no AI-generated text.

*Licence:* CC BY 4.0 (data/acl_2017/LICENSE.txt and data/conll_2016/LICENSE.txt in the release: the Creative Commons Attribution 4.0 International legal code, covering each section's reviews and paper drafts, which the reviewers and authors opted in for this dataset). Credit PeerRead (Kang et al., NAACL 2018) and the ACL 2017 / CoNLL 2016 reviewers and authors. The repository itself has no top-level licence; the iclr_2017 and arXiv sections carry none and are not included.

### `pg19`

PG-19 Language Modelling Benchmark. Rae, Potapenko, Jayakumar, Hillier & Lillicrap (ICLR 2020).

Every book of every split of the official bucket: train 28,602, validation 50, test 100 (28,752; the README's counts). One row per book, splits in the order train, validation, test and books by numeric id. `text` = the book file as released (DeepMind already removed Project Gutenberg's licence header and footer and replaced words on Ofcom's list of offensive discriminatory words with placeholder tokens <DWnn>; the README says these are the only changes). `title` = metadata.csv `short_book_title` (title and author as one string, e.g. 'A Voyage to the Moon by George Tucker'); `date` = metadata.csv `publication_date` (year; PG-19 books were published before 1919, a few dates are noisy); `source_url` = metadata.csv `url` (the Project Gutenberg ebook page). `source_record_id` = the book id, which is the Project Gutenberg ebook number: PG-19 is a subset of Project Gutenberg, and 28,674 of its books also appear in the project_gutenberg subset (in the feed's current transcription, without the <DWnn> replacements; 77 are no longer in the feed and 1 is marked copyrighted there); de-duplicate by `meta.gutenberg_id` (= project_gutenberg `source_record_id`). `meta` = metadata.csv row (book_id, short_book_title, publication_date, url), gutenberg_id, and the bucket object (file, size, md5). No author ids. Nothing is left out.

*Licence:* PG-19 is released under the Apache License 2.0 (LICENSE in the bucket and the GitHub repo); the book texts are from Project Gutenberg and in the public domain in the USA (published before 1919). Rows carry apache-2.0, the release licence (keep its notice and attribution).

### `prism`

The PRISM Alignment Dataset. Kirk, Whitefield, Röttger, Bean, Margatina, Ciro, Mosquera, Bartolo, Williams, He, Vidgen & Hale (NeurIPS 2024 Datasets and Benchmarks).

One row per conversation, all 8,011 (`conversations.jsonl`; no official split, so `full`). Every language and every flagged conversation is kept. Participants chose a conversation type (`category`: unguided, values guided, controversy guided) and wrote an opening prompt; four models answered the first turn, the participant rated them, and the conversation went on with the best-rated model, which gave two answers per later turn (2-22 turns). `messages` holds the participant's turns in order, each followed by one assistant turn with `content` null (the chosen answer's place): model responses are CC BY-NC 4.0 while these rows are CC BY 4.0 (human-written text). `model` is null because several models answered. `meta` keeps the conversation's other fields as released (`conversation_id`, `user_id`, `conversation_type`, `conversation_turns`, ratings in `performance_attributes` / `choice_attributes`, the participant's `open_feedback`, `generated_datetime`, timings, `included_in_balanced_subset`); `conversation_history` is kept without any `content` (user turns are in `messages`, model text is left out), so it still lists every model, provider, score and choice per turn; `opening_prompt` equals the first user turn (checked) and is not repeated. Joined from the other files of the release: `meta.metadata` holds the release's language, PII and moderation flags of each user turn (`user_prompts`, in turn order, via `utterances.jsonl`) and of the open feedback (`metadata.jsonl`; flags of model responses are left out with the responses), and `meta.survey` the participant's full survey record (`survey.jsonl`: demographics, location, LLM familiarity and use, stated preferences, and the free-text `self_description` and `system_string`). The 104 surveyed participants who had no conversation have no row. `utterances.jsonl` holds the same conversations in long form and adds nothing else. `date` is `generated_datetime` (2023-11-22 to 2023-12-22).

*Licence:* Dataset card: "Human-written texts (including prompts) within the dataset are licensed under CC-BY-4.0. Model responses are licensed under CC-BY-NC-4.0" (card tag `cc`); model responses are not included here. Terms of use: research and educational use; users agree not to try to re-identify or de-anonymise any individual. May contain unsafe or offensive text.

### `project_gutenberg`

Project Gutenberg (bulk plain-text feed). Project Gutenberg Literary Archive Foundation, Project Gutenberg ebooks (bulk feed snapshot 2026-10-04).

Every ebook of the bulk plain-text feed (78,004 files cache/epub/<n>/pg<n>.txt, all languages, all UTF-8), one row per ebook in ebook-number order, split 'full', except 416 ebooks whose RDF record says 'Copyrighted. Read the copyright notice inside this book for details.' (copyrighted works Project Gutenberg distributes with the holder's permission; they are not public domain and carry no licence of this corpus, so they are left out). Every other ebook's RDF record says 'Public domain in the USA.'. 77,588 rows. `text` = the ebook file between the line '*** START OF THE (or THIS) PROJECT GUTENBERG EBOOK ...' and the line '*** END OF THE PROJECT GUTENBERG EBOOK ...', both lines excluded (every file has exactly one END line; two files, 49935 and 67098, carry a second copy of the licence header and START line, and their text starts after the last START line). So the Project Gutenberg licence header and footer are removed, as the Project Gutenberg License requires when a work is redistributed without the Project Gutenberg trademark. The text between the markers is kept as released, including producer credits ('Produced by ...'), tables of contents, transcriber's notes, footnotes and illustration tags. `source_record_id` = the ebook number; `source_url` = https://www.gutenberg.org/ebooks/<n>; `title` = catalogue Title; `date` = catalogue Issued, the ebook's release date on Project Gutenberg (not the date the work was written or first published, so ebooks released after 2022-11-30 have pre_chatgpt = false although the works are old); for the 8 ebooks missing from this catalogue snapshot (76962, 77077, 77083, 77166, 77198, 78455, 79721, 79722) title and issued date come from their RDF record and meta.catalog = null. `author_id` = Project Gutenberg's agent id(s) of the RDF dcterms:creator(s) ('pg-agent:<id>', ';'-joined if several; empty if none). `meta` = the ebook's catalogue row as released (Type, Issued, Title, Language, Authors, Subjects, LoCC, Bookshelves), the RDF rights statement and creators [{agent, name}], and the file name and size. The RDF records' other fields (including the automatically generated summaries in marc520) are not used.

*Licence:* Public domain in the USA (each ebook's RDF rights statement 'Public domain in the USA.'); copyright status outside the USA is not checked and differs by country. The texts are used without the Project Gutenberg trademark: the licence header and footer are removed, so this corpus is not a Project Gutenberg distribution. Catalogue data: public domain.

### `promptevals`

PromptEvals: A Dataset of Assertions and Guardrails for Custom Production LLM Pipelines. Vir, Shankar, Chase, Fu-Hinthorn & Parameswaran (NAACL 2025).

Every row of the three official splits: train.csv (1,252), validation.csv (418), test.csv (419). `prompt` = the `Prompt Template` as released (a developer's pipeline prompt with {placeholders}, often prefixed by its LangChain message type such as 'HumanMessagePromptTemplate'); it is the single user turn. `meta` keeps the split file's other fields unchanged: ID, Domain 1-3 (`category` = Domain 1), Ground Truth Constraints (JSON list of {constraint, category}) and, in train and validation, Ground Truth Constraints (Numbered). The release's dataset.csv is the union of the three splits (the same 2,087 IDs, templates and domains), so its rows are not repeated; its own criteria column (`Ground Truth Criteria`, numbered text, which differs from the split files' numbered constraints for 22 rows) is added to `meta`. One ID (rag-prompt-642e7583-...) occurs in all three splits with the same template; its validation and test copies get ids ending in `#2` and `#3`. One validation row has an empty template and is dropped by the builder. The templates come from the LangChain Prompt Hub (snapshot May 2024); developers consented to publication (dataset card). No per-row dates.

*Licence:* MIT (dataset card). Developers can ask for their prompts to be deleted; the authors update the dataset to honour such requests.

### `rcv1_reuter_50_50`

Reuter_50_50 (RCV1 authorship subset). Liu, Reuter_50_50 (UCI Machine Learning Repository 2011, doi:10.24432/C5DS42); Lewis, Yang, Rose & Li (JMLR 2004) for RCV1.

Every story of both official parts: C50train 2,500 -> split `train`, C50test 2,500 -> split `test` (50 authors x 50 stories each). One row per file <part>/<author>/<itemid>newsML.txt, in zip order. `text` = the file (UTF-8; one paragraph per line) with the release's escaped ampersands decoded ('&amp;' -> '&': the files escape every '&' and contain no other entity). `author_id` = the author directory name (e.g. 'AaronPressman'); `source_record_id` = the RCV1 itemid from the file name (unique over both parts). The files carry no title or date: `date` = 1997, an upper bound (RCV1 holds Reuters stories of 1996-08-20 .. 1997-08-19), not the story's own year. `meta` = part, author, itemid, file. Nothing is left out. The full RCV1 (about 806,000 stories, NIST agreement) is not part of this release.

*Licence:* CC BY 4.0, as stated by the UCI Machine Learning Repository for this dataset (attribution: cite the UCI dataset, doi:10.24432/C5DS42). The stories are Reuters journalism from RCV1, which Reuters makes available for research under NIST agreements.

### `realguardrails_distractors`

RealGuardrails / SystemCheck: A Closer Look at System Prompt Robustness. Mu, Lu, Lavery & Wagner (arXiv 2025).

Three configs of the release, each a subset, every row. `realguardrails_prompts` (config `prompts`, 3,082): system prompts scraped from the ChatGPT GPT store (606, source `gpts`, gathered by the LouisShark/chatgpt_system_prompt and 0xeb/TheBigPromptLibrary projects) and HuggingChat assistants (2,476, newly scraped), deduplicated and English only. `system` = the system prompt (`instructions`). These rows have no user turn, but the record type needs `prompt`, so `prompt` repeats the system prompt and `messages` is null. `meta` keeps title, description, source, clauses (the guardrail clauses the authors extracted) and the is_test / is_sft / is_dpo flags. `date` = the creation time in a HuggingChat assistant's id (a MongoDB ObjectId, 2024-01-26 to 2024-05-23); GPT ids carry no date. `realguardrails_handwritten` (config `handwritten`, 239): user messages written by the authors against the 14 test-only system prompts; the multi-turn ones (53) contain assistant turns from the authors' manual sessions with Gemini 1.5 (README), kept as released. `realguardrails_distractors` (config `distractors`, 504): procedurally built test conversations whose earlier turns are in-context demonstrations of an unrelated task (translation pairs from facebook/flores, roleplay from google/Synthetic-Persona-Chat), either as alternating user/assistant turns (is_multiturn) or packed into one user message, followed by the test query. In both test subsets the system message goes to `system`, the remaining turns to `messages`, and `meta` keeps guardrails (and banlist / distractor_task / num_turns / is_multiturn) as released. `prompt` is the first user turn as everywhere, which in multi-turn distractor rows is the first demonstration; the test query is the last user turn. `id` there is the system prompt's id, so `source_record_id` = `<id>-<n>` (n-th row with that id, in file order). Split `full` (no splits in the release). Not included: the configs train_sft (18,497) and train_dpo (9,968), whose user messages were generated by Claude 3.5 Sonnet and responses by GPT-4o / Mistral 7B (model-written training data, not prompts people wrote), and systemmix, simplemix and preferencemix, training mixtures that add other datasets (Multifaceted, Glaive, SPML, Tulu 3, SlimOrca) under their own licences.

*Licence:* MIT (dataset card and GitHub LICENSE). Distractor rows embed third-party demonstrations: translation rows carry sentences of FLORES-200 (CC BY-SA 4.0) and are licensed here as CC BY-SA 4.0; roleplay rows carry Synthetic-Persona-Chat dialogues (CC BY 4.0) and are licensed CC BY 4.0 (the MIT parts are compatible with both). The ChatGPT-store system prompts were collected by third-party projects from GPTs whose creators wrote them.

### `realguardrails_handwritten`

RealGuardrails / SystemCheck: A Closer Look at System Prompt Robustness. Mu, Lu, Lavery & Wagner (arXiv 2025).

Three configs of the release, each a subset, every row. `realguardrails_prompts` (config `prompts`, 3,082): system prompts scraped from the ChatGPT GPT store (606, source `gpts`, gathered by the LouisShark/chatgpt_system_prompt and 0xeb/TheBigPromptLibrary projects) and HuggingChat assistants (2,476, newly scraped), deduplicated and English only. `system` = the system prompt (`instructions`). These rows have no user turn, but the record type needs `prompt`, so `prompt` repeats the system prompt and `messages` is null. `meta` keeps title, description, source, clauses (the guardrail clauses the authors extracted) and the is_test / is_sft / is_dpo flags. `date` = the creation time in a HuggingChat assistant's id (a MongoDB ObjectId, 2024-01-26 to 2024-05-23); GPT ids carry no date. `realguardrails_handwritten` (config `handwritten`, 239): user messages written by the authors against the 14 test-only system prompts; the multi-turn ones (53) contain assistant turns from the authors' manual sessions with Gemini 1.5 (README), kept as released. `realguardrails_distractors` (config `distractors`, 504): procedurally built test conversations whose earlier turns are in-context demonstrations of an unrelated task (translation pairs from facebook/flores, roleplay from google/Synthetic-Persona-Chat), either as alternating user/assistant turns (is_multiturn) or packed into one user message, followed by the test query. In both test subsets the system message goes to `system`, the remaining turns to `messages`, and `meta` keeps guardrails (and banlist / distractor_task / num_turns / is_multiturn) as released. `prompt` is the first user turn as everywhere, which in multi-turn distractor rows is the first demonstration; the test query is the last user turn. `id` there is the system prompt's id, so `source_record_id` = `<id>-<n>` (n-th row with that id, in file order). Split `full` (no splits in the release). Not included: the configs train_sft (18,497) and train_dpo (9,968), whose user messages were generated by Claude 3.5 Sonnet and responses by GPT-4o / Mistral 7B (model-written training data, not prompts people wrote), and systemmix, simplemix and preferencemix, training mixtures that add other datasets (Multifaceted, Glaive, SPML, Tulu 3, SlimOrca) under their own licences.

*Licence:* MIT (dataset card and GitHub LICENSE). Distractor rows embed third-party demonstrations: translation rows carry sentences of FLORES-200 (CC BY-SA 4.0) and are licensed here as CC BY-SA 4.0; roleplay rows carry Synthetic-Persona-Chat dialogues (CC BY 4.0) and are licensed CC BY 4.0 (the MIT parts are compatible with both). The ChatGPT-store system prompts were collected by third-party projects from GPTs whose creators wrote them.

### `realguardrails_prompts`

RealGuardrails / SystemCheck: A Closer Look at System Prompt Robustness. Mu, Lu, Lavery & Wagner (arXiv 2025).

Three configs of the release, each a subset, every row. `realguardrails_prompts` (config `prompts`, 3,082): system prompts scraped from the ChatGPT GPT store (606, source `gpts`, gathered by the LouisShark/chatgpt_system_prompt and 0xeb/TheBigPromptLibrary projects) and HuggingChat assistants (2,476, newly scraped), deduplicated and English only. `system` = the system prompt (`instructions`). These rows have no user turn, but the record type needs `prompt`, so `prompt` repeats the system prompt and `messages` is null. `meta` keeps title, description, source, clauses (the guardrail clauses the authors extracted) and the is_test / is_sft / is_dpo flags. `date` = the creation time in a HuggingChat assistant's id (a MongoDB ObjectId, 2024-01-26 to 2024-05-23); GPT ids carry no date. `realguardrails_handwritten` (config `handwritten`, 239): user messages written by the authors against the 14 test-only system prompts; the multi-turn ones (53) contain assistant turns from the authors' manual sessions with Gemini 1.5 (README), kept as released. `realguardrails_distractors` (config `distractors`, 504): procedurally built test conversations whose earlier turns are in-context demonstrations of an unrelated task (translation pairs from facebook/flores, roleplay from google/Synthetic-Persona-Chat), either as alternating user/assistant turns (is_multiturn) or packed into one user message, followed by the test query. In both test subsets the system message goes to `system`, the remaining turns to `messages`, and `meta` keeps guardrails (and banlist / distractor_task / num_turns / is_multiturn) as released. `prompt` is the first user turn as everywhere, which in multi-turn distractor rows is the first demonstration; the test query is the last user turn. `id` there is the system prompt's id, so `source_record_id` = `<id>-<n>` (n-th row with that id, in file order). Split `full` (no splits in the release). Not included: the configs train_sft (18,497) and train_dpo (9,968), whose user messages were generated by Claude 3.5 Sonnet and responses by GPT-4o / Mistral 7B (model-written training data, not prompts people wrote), and systemmix, simplemix and preferencemix, training mixtures that add other datasets (Multifaceted, Glaive, SPML, Tulu 3, SlimOrca) under their own licences.

*Licence:* MIT (dataset card and GitHub LICENSE). Distractor rows embed third-party demonstrations: translation rows carry sentences of FLORES-200 (CC BY-SA 4.0) and are licensed here as CC BY-SA 4.0; roleplay rows carry Synthetic-Persona-Chat dialogues (CC BY 4.0) and are licensed CC BY 4.0 (the MIT parts are compatible with both). The ChatGPT-store system prompts were collected by third-party projects from GPTs whose creators wrote them.

### `reddit_tifu`

Reddit TIFU. Kim, Kim & Kim, Abstractive Summarization of Reddit Posts with Multi-level Memory Networks (NAACL 2019).

Every line of the release file (79,949 posts; no official split). One row per line: `text` = `selftext` (the post as posted, original case, with its TL;DR line where it has one; kept as released: Reddit's HTML escapes are already decoded in this field, and the 39 posts that still contain '&amp;', '&gt;' or '&lt;' are left so), `title` = `title`, `date` = `created_utc` (UTC), `source_url` = `url`, `source_record_id` = `id` (the Reddit post id; post 2pqihw occurs twice, as two crawls differing only in score and ups: the second occurrence, line 55,410, gets '2pqihw@line55410'). `meta` holds every other released field unchanged: `tldr` (TIFU-long summary, lower-cased by the release; null for the 36,965 posts without a TL;DR), `trimmed_title` (TIFU-short summary: the title without 'TIFU by', lower-cased), `selftext_without_tldr` (the release's lower-cased document text), the tokenized fields (`title_tokenized`, `trimmed_title_tokenized`, `tldr_tokenized`, `selftext_without_tldr_tokenized`), `selftext_html`, `permalink`, `created_utc`, `score`, `ups`, `upvote_ratio`, `num_comments`, plus `line` (0-based line number in the file). Both versions are therefore covered: TIFU-short = every row with its title, TIFU-long = the rows with a non-null meta.tldr. No author names are released. The 199 posts whose `selftext` is empty (title-only) are counted by the builder as dropped_empty. Nothing else is left out.

*Licence:* MIT (github.com/ctr4si/MMN README, 'License: MIT license'). The posts are Reddit user content.

### `slopbench`

SlopBench: How Well Can We Rank Language Models by Slop?. Roongta, Gaddipati & Huynh (arXiv 2026).

All 112 scenarios of scenarios/pilot/{email,social,essay,slack}.json, in both of the release's splits (`public` 74, `heldout` 38; the scenario's own `split` field). Each scenario is one prompt as the benchmark sends it (harness/generate.py): `system` = the domain's scaffold and the single user turn = build_user_prompt(), i.e. 'About me: <persona>', 'Who this is going to: <recipient>', the prior thread messages ('They wrote: ...' / 'I previously wrote: ...') and the request itself; this text was checked to equal the `user_prompt` recorded for that scenario in the release's runs/full-merged/outputs.jsonl. `meta` keeps every scenario field unchanged (including the bare request `prompt`, persona, recipient, context_thread, goal, length_target, author_note, pilot), the one-line setup from scenarios/pilot/setups.json and the source file. `category` = the scenario category (e.g. cold_outreach); the domain is in `meta.domain`. The prompts were written by the authors in 2026. Not included: the models' generated outputs in runs/ (model text, not prompts).

*Licence:* MIT (repository LICENSE, Copyright (c) 2026 Slashy; docs/DATASET_SPEC.md: 'scenarios ... -> MIT'). The paper also states that the prompts are released under CC BY 4.0. Attribute SlopBench.

### `stack_exchange`

Stack Exchange Data Dump (2024-04-02). Stack Exchange, Inc. (Stack Exchange Data Dump, Internet Archive, 2024-04-02 release).

Every row of Posts.xml of every site archive in the dump (363 sites, main and meta) and of stackoverflow.com-Posts.7z: questions, answers and the other post types (tag wikis and excerpts, moderator nominations, wiki placeholders, privilege pages). Deleted posts are not in the public dump. The other files (comments, users, votes, badges, post history, post links, tags) are not used; Stack Overflow's other files were not downloaded. `text` = Body, which is HTML (the latest revision), converted to plain text much as a browser shows it: paragraphs, code blocks, lists, block quotes, headings and tables set off by a blank line; list items and table rows on their own lines; `<br>` a line break; HTML whitespace collapsed outside code; code blocks (`<pre>`) kept exactly as written (indentation, blank lines; newlines at their ends trimmed); inline markup (code, links, emphasis) kept as its text, without link targets; table cells separated by a tab; images left out (counted in meta); outside code at most one blank line in a row; no list numbers or quote markers added. Nothing else is dropped or filtered; rows whose Body is empty (mostly empty tag wikis) are counted by the builder. `title` = Title (questions); `prompt` is not filled (the question of an answer is linked by `parent_id`); `parent_id` = '<site>:<ParentId>' for answers and '<site>:<Id>' for questions; `date` = CreationDate (the text may have been edited later: LastEditDate in meta); `author_id` = '<site>:<OwnerUserId>' (-1 is the Community bot); `source_record_id` = '<site>:<Id>'; `source_url` = https://<site>/questions/<Id> or https://<site>/a/<Id> (none for other post types); `genre` from PostTypeId (Q&A question, Q&A answer, tag wiki, tag wiki excerpt, orphaned tag wiki, moderator nomination, wiki placeholder, privilege wiki). `meta` = the site and every other Posts.xml attribute as released (strings: Id, PostTypeId, ParentId, AcceptedAnswerId, Score, ViewCount, Tags, OwnerUserId, OwnerDisplayName, LastEditorUserId, LastEditorDisplayName, CreationDate, LastEditDate, LastActivityDate, CommunityOwnedDate, ClosedDate, ContentLicense, AnswerCount, CommentCount, FavoriteCount, ...) plus counts from the conversion: `body_html_chars` (length of the HTML Body), `n_code_blocks`, `code_chars`, `n_images`, `quote_chars` (characters inside `<blockquote>`). Order: sites by name; within a site, Posts.xml order.

*Licence:* CC BY-SA, version per post (ContentLicense: 2.5, 3.0 or 4.0), as stated in the dump's license.txt and the archive.org item. Attribution: name the site, link the post and its author; share-alike applies to adaptations.

### `supreme_court_oral_arguments`

U.S. Supreme Court oral argument transcripts (OT2000-OT2026). Supreme Court of the United States, oral argument transcripts (supremecourt.gov), October Terms 2000-2026.

Every transcript PDF listed on the Court's argument-transcript pages for October Terms 2000-2026 as of 2026-10-13: 1,809 transcripts (arguments 2000-10-02 to 2026-10-07), one row each (a PDF that covers several consolidated dockets is one row; the listing's dockets are in meta.dockets). Earlier terms are not on supremecourt.gov. Text comes from the PDF's text layer (pypdf 5.1.0). Page furniture is removed: the line numbers 1-25, page numbers, the 'Official - Subject to Final Review' banner and the court reporter's footer (Alderson / Heritage Reporting). `text` = the proceedings, from the 'P R O C E E D I N G S' marker to the closing '(Whereupon, ... the case was submitted.)', as paragraphs separated by blank lines: one paragraph per speaker turn written 'SPEAKER TAG: words' (tag as printed, e.g. 'CHIEF JUSTICE ROBERTS', 'JUSTICE KAGAN', 'MR. SMITH'; 'QUESTION' = an unnamed justice, the Court's practice before OT2004), its hard-wrapped lines joined with single spaces (a word hyphenated at a line end is re-joined), one paragraph per section heading ('ORAL ARGUMENT OF X ON BEHALF OF THE PETITIONER', 'REBUTTAL ARGUMENT OF ...') and per untagged line run (e.g. the opening time). Stage directions such as '(Laughter.)' stay as printed. The cover, appearances and contents pages (the lines before the marker) are in meta.front_matter, one printed line per line; the alphabetical word index at the end of each transcript (a page:line concordance, not text) is left out. meta.blocks lists the paragraphs of `text` in order (paragraph k of text split on blank lines = block k) with kind (turn / heading / untagged / closing), speaker tag, role (justice / counsel / other, from the tag) and segment (preliminary, argument_opening, argument_qa, rebuttal_opening, rebuttal_qa). A transcript whose turns cannot be found is kept with all its cleaned lines as text and meta.parsed = false. `title` = case name from the listing; `date` = the listing's Date Argued (the cover date when the listing's date is malformed); `source_record_id` = '<term>/<PDF file stem>'; `source_url` = the PDF; `genre` 'U.S. Supreme Court oral argument transcript'. meta also holds term, dockets, date_argued_listing, cover_date, pdf_sha256, pdf_bytes, pdf_last_modified, n_pages. Arguments from 2020-05-04 to 2021-05 were held by telephone.

*Licence:* No licence stated. Official U.S. Supreme Court record published free to the public on supremecourt.gov ('Same-day transcripts are considered official but subject to final review'); prepared by the Court's contracted official reporter (Alderson Reporting Company to OT2016, Heritage Reporting Corporation from OT2017). Treated as a public-domain U.S. government record.

### `un_general_debates`

UN General Debate Corpus (UNGDC) 1970-2018. Baturo, Dasandi & Mikhaylov (Research & Politics 2017).

Every statement of the release: the 8,093 text files <ISO3>_<session>_<year>.txt of UNGDC_1970-2018.zip (sessions 25-73, 1970-2018), one row each. Left out of the zip: macOS/RStudio junk files (.DS_Store, .Rhistory, .Rapp.history, '.DS_Store-to-UTF-8.txt', __MACOSX). `text` = the file as released (the official UN English text of the statement; UTF-8, the byte-order mark dropped); statements of sessions 25-47 (1970-1992) were OCR'd from typewritten records by the authors and keep OCR errors. `source_record_id` = file stem; `date` = the session year (the General Debate is held in September-October); `author_id` = the ISO3 code of the speaking country (the release's code, e.g. 'EU' for the European Union); `genre` 'UN General Debate statement'. `meta`: iso_code, session, year, zip_member, and `speakers` = the rows of posts_of_speakers.xlsx (sheet 1: Year, ISO Code, Country, Name of Person Speaking, Post, Language; it covers 1970-2014) with the same ISO code and year, as released (none for 2015-2018). Newer version not used: the authors' current release is Harvard Dataverse doi:10.7910/DVN/0TJX8Y v14.0 (2026-03-08; UNGDC_1946-2025.tar.gz, 11,141 statements 1946-2025, CC0 1.0), but its files can only be downloaded after filling in a Dataverse guestbook with a name, email, institution, position and purpose (API answer: 'You may not download this file without the required Guestbook response'); the GitHub release is the latest version available without that form. The GitHub repo's HEAD (86c8a39, 2019-04-02) removed the zip and points to Dataverse; the pinned commit is the last one holding it.

*Licence:* CC0 1.0: the licence of the authors' Dataverse deposit of the corpus (the GitHub repo states no licence). The texts are official records of the UN General Assembly.

### `un_parallel`

The United Nations Parallel Corpus v1.0. Ziemski, Junczys-Dowmunt & Pouliquen (LREC 2016).

Subset un_parallel (split full): every document of the TEI XML release, one row per document and language: 799,276 TEI files (ar 116,751, en 159,336, es 130,275, fr 158,800, ru 138,168, zh 95,946), parliamentary documents and official records from the UN Official Document System (ODS) published 1990-2014 (year folders: en, fr, ru from 1990, es 1991, ar 1992, zh 1994), in the six official languages; most language versions are human translations of one original. A TEI file holds the document as paragraphs <p> of automatically split sentences <s>; `text` = the paragraphs joined by newlines, the sentences of a paragraph joined by a space (by nothing in Chinese, which writes no space between sentences); sentence text as released, including the release's conversion artefacts (running heads, page and job numbers, symbols on their own line; in some early Arabic documents presentation-form letters and digits in reverse order). 103 documents have no text and are dropped by the builder: 15 zero-byte files (A/RES/49/34-37 in ar, en, es, fr, ru) and 88 files without any sentence text; 799,173 rows remain. 10 files are not well-formed XML because the symbol holds a bare '&' (e.g. E/C.12/Q/T&T/1): the ampersand is escaped before parsing (meta.xml_repair). Translations share meta.parallel_id = '<year>/<symbol path>', the release's own alignment key (it puts a document's language versions at the same path in each language folder; the path is the document symbol in file-name form, e.g. 2014/a/c_3/69/l_1/rev_1 for A/C.3/69/L.1/REV.1, which is in meta.symbol). The symbol itself is not used as the key because 171 files (a few dozen documents whose symbol holds '(', a space, a range or several symbols, e.g. A/52/7(CHAP.II)..., A/68/40 (VOL.II), A/C.5/51/SR.1-54) have a malformed header in which the symbol is cut off and the other fields are shifted (e.g. the date field holds the job number), so different documents would share it. 163,814 parallel ids: 92,313 in all six languages, 7 in one only. `source_record_id` = '<lang>/<parallel_id>'; `date` = the publication date of the TEI header (YYYYMMDD as ISO), for the 171 malformed headers the year folder (meta.date_from = 'year folder'); `source_url` = the document's ODS link https://docs.un.org/<lang>/<symbol> (empty for malformed headers); meta: symbol, jobno (UN job number), pub_place, date as released, keywords (the subject terms of the header, when given), file_language, path, n_paragraphs, n_sentences. Subset un_parallel_testsets (splits dev, test): the release's official development and test sets (UNv1.0.testsets.tar.gz): 4,000 sentences each, aligned 1-1 across the six languages, drawn from documents released in the first (dev) and second (test) quarter of 2015, which are not in the TEI release; one row per sentence per language (48,000 rows), meta.parallel_id = '<split>/<line>', `parent_id` = the document path given in the .ids file, date 2015. Left out, because they repeat the TEI text: the 15 pairwise plain-text bitexts (UNv1.0.<l1>-<l2>.tar.gz.NN) and the fully aligned 6-way sentence corpus (UNv1.0.6way.tar.gz.NN), both built from the same documents; also the TEI link files (UNv1.0-TEI.links.tar.gz.NN: sentence alignments between language versions, no text) and the README, DISCLAIMER and paper PDF in the archives.

*Licence:* Official records and parliamentary documents of the United Nations, in the public domain; the release's DISCLAIMER ("no other restrictions apply") requires users to acknowledge the United Nations as the source and to cite Ziemski et al. (2016); no warranty; the privileges and immunities of the United Nations are reserved.

### `un_parallel_testsets`

The United Nations Parallel Corpus v1.0. Ziemski, Junczys-Dowmunt & Pouliquen (LREC 2016).

Subset un_parallel (split full): every document of the TEI XML release, one row per document and language: 799,276 TEI files (ar 116,751, en 159,336, es 130,275, fr 158,800, ru 138,168, zh 95,946), parliamentary documents and official records from the UN Official Document System (ODS) published 1990-2014 (year folders: en, fr, ru from 1990, es 1991, ar 1992, zh 1994), in the six official languages; most language versions are human translations of one original. A TEI file holds the document as paragraphs <p> of automatically split sentences <s>; `text` = the paragraphs joined by newlines, the sentences of a paragraph joined by a space (by nothing in Chinese, which writes no space between sentences); sentence text as released, including the release's conversion artefacts (running heads, page and job numbers, symbols on their own line; in some early Arabic documents presentation-form letters and digits in reverse order). 103 documents have no text and are dropped by the builder: 15 zero-byte files (A/RES/49/34-37 in ar, en, es, fr, ru) and 88 files without any sentence text; 799,173 rows remain. 10 files are not well-formed XML because the symbol holds a bare '&' (e.g. E/C.12/Q/T&T/1): the ampersand is escaped before parsing (meta.xml_repair). Translations share meta.parallel_id = '<year>/<symbol path>', the release's own alignment key (it puts a document's language versions at the same path in each language folder; the path is the document symbol in file-name form, e.g. 2014/a/c_3/69/l_1/rev_1 for A/C.3/69/L.1/REV.1, which is in meta.symbol). The symbol itself is not used as the key because 171 files (a few dozen documents whose symbol holds '(', a space, a range or several symbols, e.g. A/52/7(CHAP.II)..., A/68/40 (VOL.II), A/C.5/51/SR.1-54) have a malformed header in which the symbol is cut off and the other fields are shifted (e.g. the date field holds the job number), so different documents would share it. 163,814 parallel ids: 92,313 in all six languages, 7 in one only. `source_record_id` = '<lang>/<parallel_id>'; `date` = the publication date of the TEI header (YYYYMMDD as ISO), for the 171 malformed headers the year folder (meta.date_from = 'year folder'); `source_url` = the document's ODS link https://docs.un.org/<lang>/<symbol> (empty for malformed headers); meta: symbol, jobno (UN job number), pub_place, date as released, keywords (the subject terms of the header, when given), file_language, path, n_paragraphs, n_sentences. Subset un_parallel_testsets (splits dev, test): the release's official development and test sets (UNv1.0.testsets.tar.gz): 4,000 sentences each, aligned 1-1 across the six languages, drawn from documents released in the first (dev) and second (test) quarter of 2015, which are not in the TEI release; one row per sentence per language (48,000 rows), meta.parallel_id = '<split>/<line>', `parent_id` = the document path given in the .ids file, date 2015. Left out, because they repeat the TEI text: the 15 pairwise plain-text bitexts (UNv1.0.<l1>-<l2>.tar.gz.NN) and the fully aligned 6-way sentence corpus (UNv1.0.6way.tar.gz.NN), both built from the same documents; also the TEI link files (UNv1.0-TEI.links.tar.gz.NN: sentence alignments between language versions, no text) and the README, DISCLAIMER and paper PDF in the archives.

*Licence:* Official records and parliamentary documents of the United Nations, in the public domain; the release's DISCLAIMER ("no other restrictions apply") requires users to acknowledge the United Nations as the source and to cite Ziemski et al. (2016); no warranty; the privileges and immunities of the United Nations are reserved.

### `webis_tldr_17`

Webis-TLDR-17. Völske, Potthast, Syed & Stein, TL;DR: Mining Reddit to Learn Automatic Summarization (Workshop on New Frontiers in Summarization, EMNLP 2017).

Every line of the release's single JSON file (3,848,330 posts: 1,658,567 submissions, ids 't3_...', and 2,189,763 comments, merged by the authors on one schema; no official split). One row per line: `text` = `body` (the post as written, including its TL;DR line; the authors had already removed URLs from it, leaving e.g. '[link text]( '), with Reddit's HTML escaping undone: '&lt;', '&gt;' and '&amp;' (Reddit's escapes, found in about 10% of bodies, mostly '&gt;' quote markers) become '<', '>' and '&', as the author typed them; nothing else is changed. `author_id` = `author`, `source_record_id` = `id` (unique in the file). The author-written summary is `meta.summary` (the text after the TL;DR marker) and `meta.content` the text before it, both exactly as released (entities kept), with `meta.normalizedBody` (the release's whitespace-normalised body), `content_len`, `summary_len`, `subreddit`, `subreddit_id` and, where present, `title`, all unchanged, plus `line` (0-based line number). `title` = the title of a submission; 104,462 comments also carry a `title` in the release (apparently the title of their thread), which stays in meta.title only. `genre` = 'Reddit submission' or 'Reddit comment' (by the 't3_' id prefix). `source_url` = https://www.reddit.com/comments/<id>/ for submissions, None for comments (the release has no thread id for them). The release has no timestamps: `date` = '2016', an upper bound (the posts are from 2006-2016), not the year of writing. Nothing is left out.

*Licence:* CC BY 4.0 (Zenodo record 1043504). The posts are Reddit user content and remain subject to Reddit's terms.

### `wikiconv`

WikiConv (English). Hua, Danescu-Niculescu-Mizil, Taraborelli, Thain, Sorensen & Dixon, WikiConv: A Corpus of the Complete Conversational History of a Large Online Collaborative Community (EMNLP 2018).

Every line of all 500 English NDJSON files (241,288,668 lines, one per conversational action; files in name order, lines in file order; 210,018,590 rows after the 31,270,078 actions with empty cleaned text). One row per action: CREATION (new section heading), ADDITION (new comment), MODIFICATION (new version of a comment), DELETION (comment removed: the release stores the removed text in the record) and RESTORATION (removed comment put back). `text` = `cleaned_content`, the release's plain text of the action (MediaWiki markup, templates and user signatures removed by the authors with mwparserfromhell). The release's raw wikitext of the action (its `content` field, about 40% of the release's bytes) is left out (owner decision, to keep the subset small); it is in the release files (gs://wikidetox-wikiconv-public-dataset/dataset/English/, field `content`, matched by `id`, or by meta.shard and meta.line). Actions whose cleaned text is empty (markup only: project banners, templates, images, a bare signature) are counted by the builder as dropped_empty; nothing else is left out, and DELETION and RESTORATION records are kept although their text repeats text of an earlier action (filter on meta.type to leave them out). `source_record_id` = `id` (the release's action id '<rev_id>.<offset>.<offset>'); `author_id` = `authors` as released ('<user id>:<user name>', or 'ANONYMOUS:<IP address>'): the user the release credits with the text. For ADDITION and CREATION this is the user who posted it; for MODIFICATION, DELETION and RESTORATION it is often another user than the one who made the edit, removal or restoration (that user is meta.user_text / meta.user_id; in shards 0, 250 and 499 they differ for 40% of MODIFICATION, 89% of DELETION and 75% of RESTORATION records); `parent_id` = `conversation_id` (the id of the action that started the conversation, usually its CREATION row: rows of one discussion thread share it); `date` = `timestamp` (UTC, the revision time of the action); `source_url` = the revision diff https://en.wikipedia.org/w/index.php?oldid=<rev_id>&diff=prev; `title` = None (the talk page is meta.page_title); `genre` from `type`: section heading, comment, comment (modified version), comment (removed text), comment (restored text). `meta` = every other field of the record unchanged except `content`: type, page_title, page_id, rev_id, conversation_id, parent_id (the action a MODIFICATION, DELETION or RESTORATION acts on), replyTo_id, ancestor_id, indentation, user_text, user_id, timestamp, isUnchanged, toxicity and sever_toxicity (the release's Perspective API scores, field names as released), plus `shard` (file number 0-499) and `line` (0-based line in the file). Only the English release is used: the bucket's Chinese, German, Greek and Russian releases are separate corpora not requested for this subset.

*Licence:* Comment content CC BY-SA 3.0, metadata CC0 1.0 (bucket README.txt: 'The meta-data in this corpus is goverened by the CC0 license v1.0, and the content of the comments is goverened by the CC-SA license v3.0'); rows carry the stricter CC BY-SA 3.0. Attribution: Wikipedia contributors (the page and revision are in meta).

### `wikipedia`

English Wikipedia (Wikimedia dump enwiki-20220501). Wikimedia Foundation, English Wikipedia database dump of 2022-05-01 (pages-articles-multistream).

Every article of the dump: every page in namespace 0 that is not a redirect (pages with a <redirect> element are left out, as are all other namespaces: talk, user, Wikipedia, file, template, category, portal, draft, module, ...), one row per page with the page's latest revision as of the dump; split `full`. The whole dump is read: the recombined multistream file, which holds the same pages as the run's 62 split files (the earlier build used 7 of those); nothing is sampled or filtered. Disambiguation, list and set-index pages are articles in namespace 0 and are included. That is 6,492,495 pages; 4,814 of them have no text left after the conversion (e.g. lists that are one table, pages made only of templates) and are counted by the builder as dropped_empty, leaving 6,487,681 rows. `text` = plain text converted from the revision's wikitext with mwparserfromhell 0.7.2 (gen/sources/wikipedia_text.py). Templates cannot be expanded (the dump has no template definitions), so the rules are: (1) before parsing, comments (<!-- .. -->) are removed, then <ref>..</ref> and <ref ../> with their content (a ref tag as MediaWiki reads it: '<ref' followed by a space, '/' or '>'; an unclosed <ref ..> runs to the next </ref>), and the ':' indent before a table start ':{|' is removed so the table is recognised; the parser runs with skip_style_tags, and bold/italic apostrophes are removed afterwards (runs of 2, 3 or 5; of 4 one apostrophe stays; beyond 5 the extra ones stay). (2) Kept: text; HTML entities decoded; internal links as their shown text (the label, else the target as written); external links as their label (bracketed links without a label are removed, bare URLs kept); headings; the contents of formatting and other HTML tags (b, i, span, small, sub, sup, abbr, code, ...); <nowiki> content literally; <pre>, <syntaxhighlight>, <source> and <poem> content with its line breaks; <br> as a line break; <blockquote>, <div>, <center>, <p> and HTML lists as blocks of their own (HTML <li> items as '- ' items). (3) Removed with their content: every template except the rendered ones below (infoboxes, navboxes, hatnotes, citation and footnote templates, maintenance tags, {{coord}}, {{IPA..}}, ...); tables ({| |} and <table>); <references>, <gallery>, <timeline>, <graph>, <mapframe>, <imagemap>, <templatestyles>, <math>, <chem>, <ce>, <score>, <hiero> and similar non-text tags; comments; file/image/media links (with their captions); category links (their names go to meta.categories); interlanguage links ([[xx:..]] without a label); magic words (__NOTOC__ ...). Template or table markup that the parser could not parse (malformed or opened by a template) is removed afterwards: remaining '{{..}}' spans; a line that starts with an unclosed '{{' (a block template opener such as '{{Infobox settlement'; an unclosed '{{' inside a line of prose is kept, as the page shows it); and lines that start with table syntax (from '{|' through the next '|}' or heading; '|}', '|-', '|+', '|..', or '!' with '!!' or '|'). (4) Rendered templates (output as on the page, or as noted): formatting wrappers whose output is their argument (nowrap, small, big, abbr, sic, mvar, math, smallcaps, nobold, noitalic, sortname, flag (the country name, without the flag icon), ...); {{lang|code|text}} -> text, {{lang-xx|text}} -> text (the page also shows a language label, left out), {{transl|..|text}} -> text; {{ill}}/{{interlanguage link}} -> the English label; {{ship|prefix|name}} -> 'prefix name'; {{nihongo|English|kanji|romaji}} -> 'English (kanji, romaji)'; {{chem|..}} -> its arguments joined; {{ISBN|x}} -> 'ISBN x'; {{convert|v|unit}} / {{cvt}} -> 'v unit' and {{convert|v1|-|v2|unit}} -> 'v1–v2 unit' (also to/and/or/by/x ranges): the input value and unit only, without the conversion the page shows, with a few unit codes printed as their symbol (C -> °C, km2 -> km², sqmi -> sq mi, acre -> acres, ...), other codes as written; {{as of|y|m}} -> 'As of Month y'; {{circa|y}} -> 'c. y'; dash, space and punctuation templates ({{ndash}}, {{mdash}}, {{snd}}, {{nbsp}}, {{'}}, {{'s}}, {{!}}, {{=}}, {{dot}}, ...) -> their character. (5) Layout: consecutive lines of prose form one paragraph (joined with a space, as MediaWiki does); paragraphs, headings and lists are separated by a blank line; a heading is a line of its own holding the heading text (no '==' markers); list items are lines of their own: '* x' -> '- x', '# x' -> '1. x', '2. x', ... (numbered per run), ';' terms and ':' indents -> the text, nested items indented by two spaces per level; runs of spaces are collapsed; a line left with only removed material ends the paragraph. A heading whose section has no text left (e.g. 'References' holding only <references/>) is dropped. Known limits: removing a template or tag from inside a sentence can leave gaps or empty brackets ('Paris (; French: )'); meta.removed_inline counts such removals per article (templates other than footnote, citation and maintenance templates, and math/chem/score/hiero tags, removed from a line that keeps other text), so damaged articles can be found. `title` = page title; `source_record_id` = page id; `date` = the revision timestamp; `source_url` = https://en.wikipedia.org/w/index.php?oldid=<revision id> (the revision, with its history for attribution); `genre` = 'encyclopedia article'. `meta` holds page_id, revision_id, parent_revision_id, revision_sha1 (base-36 sha1 of the wikitext, as in the dump), wikitext_bytes, model, short_description (the {{Short description}} text), categories, sections ([depth, heading], depth 1 = '==' level) and removed_inline. Revision contributors and edit comments are not kept. Overlap: the `kilt` source holds most of the same articles in their 2019-08-01 revision.

*Licence:* CC BY-SA 3.0 (and GFDL): the licence of Wikipedia text at the dump date (Wikipedia moved to CC BY-SA 4.0 in June 2023; the Internet Archive item's licenseurl is creativecommons.org/licenses/by-sa/3.0). Attribution: the page history (source_url); share-alike applies.

### `wildchat`

WildChat-1M. Zhao, Ren, Hessel, Cardie, Choi & Deng (ICLR 2024).

Every conversation of the release (837,989; one split, `train`; all languages: 74 values of the release's `language`). Users talked with GPT-3.5 or GPT-4 (`model`) through a free chatbot between 2023-04-09 and 2024-04-29. ODC-BY 1.0 covers the whole conversation, so `messages` holds every user and assistant turn. `date` is the release's `timestamp` (UTC time of the conversation's last turn). `meta` keeps the release's other fields as released: `conversation_hash`, `model`, `timestamp`, `turn`, the conversation's `language`, `toxic`, `redacted`, `state`, `country`, `hashed_ip`, `header` (accept-language, user-agent), the per-utterance `openai_moderation` and `detoxify_moderation` results, and `conversation`: each utterance's own fields without its text (`role`, `language`, `toxic`, `redacted`, `turn_identifier`, `timestamp`, and for user turns `country`, `state`, `hashed_ip`, `header`), aligned with `messages`. Timestamps are ISO 8601 (UTC). `conversation_hash` occurs more than once (826,319 distinct values over 837,989 rows: identical conversations from different sessions); later copies get `#2`, `#3`, ... in release order. As the README notes, some conversations start with an empty user input; the 3,631 rows whose user turns are all empty (after normalisation) are dropped by the builder and counted. The release itself already excludes the conversations flagged as toxic (2024-07-22) and some with sensitive information (2024-10-17); the toxic ones are in the separate, gated WildChat-1M-Full, which is not included.

*Licence:* ODC-BY 1.0 (README and LICENSE.md; changed from the AI2 ImpACT licence on 2024-06-26, applied retroactively). PII was detected and anonymised by the authors (`redacted`); IP addresses are hashed. Do not try to identify users.

## Licences and attribution

Each row keeps the licence of its source (column `license`; the table above and `LICENSE.md` 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.

The columns this corpus adds (ids, language, counts, hashes, `pre_chatgpt`, genre labels) are released under CC0. The build code is Apache-2.0.

## Personal data and removal requests

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.

## Citation

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.