peterkirby/styledata

MUSE training data from STEB

Select a configuration explicitly. There is no combined/default training configuration. Labels and positive/negative sampling are scoped to each source/view. Different CORE parent views overlap.

This release preserves original document members, with outer trimming recorded. No native STEB release, frozen evaluation selection, or existing local CORE artifact was changed. The Pile and supplemental sources outside STEB are excluded. All 130 real native identifiers are accounted for in source-coverage.json; zero-training and unresolved-rights sources are explicit.

Categorical labels retain supplied ordering as metadata only. Topics and time-of-day preference are controls, never targets. Origin IDs: 0 human-only, 1 AI-only, 2 mixed. Mixed requires matched retained human wording or documented human edits; complete rewrites are AI-only. Generator labels are evidenced lineages, not vendors or size guesses. MAGE continuation prefixes use the first 30 words, or the complete shorter human document. Unresolved or ambiguous parent lineages are withheld from training; their original holdouts still reserve all possible parents. Unknown origin alone excludes only the origin view.

Relations preserve ordered, separate members and the versioned annotation schema in annotation-schemas.json. Never concatenate references, infer global authors from trial IDs, treat unrelated trials as automatic negatives, infer absolute classes from relative styles, or derive order supervision from category numbers. Generic helpfulness preferences and unsupported paraphrase pairs are excluded.

Official source splits and whole document lineages remain reserved, including category-ineligible examples. Missing validation uses source-grouped 90/10 seed 42 once across views, before category filtering. Frozen non-Pile holdouts and complete canonical probing pools take priority: test, additional original test, validation, train. Lower-priority copies are removed without refilling. Each categorical partition requires two distinct texts per class and two classes.

OneStopEnglish uses complete article files. CEECES uses complete visible letter prose, subset-specific author classes, and separately supplied sender gender/age groups; multiple or uncertain senders are excluded from sender views. X-GENRE CORE-derived documents without a recovered original lineage are quarantined. Shared paragraph wording alone never reserves an unrelated entire document.

Rights are source-specific; this collection has no blanket text license. rights.json preserves upstream attribution, grants, license text, and limitations. CORE text training/redistribution permission is user-attested, not independently verified; annotation CC-BY-SA-4.0 is separate. Noncommercial and NoDerivatives conditions remain applicable where stated. Changes consist of technical format conversion, outer trimming, annotation resolution, deduplication and exclusion. Cite and attribute the original publishers linked in rights.json.

from datasets import load_dataset
texts = load_dataset("peterkirby/styledata", "OneStopEnglishCorpus__reading_level", split="train")
relations = load_dataset("peterkirby/styledata", "StylePTB__relation_active_to_passive", split="train")
# Sample categorical positives/negatives within the selected configuration only.
# Interpret relations using annotation_json; unrelated trials are not negatives.

View	Train	Validation	Excluded/collapsed occurrences
ASSET__simplification_relation	0	1,663	696
OneStopEnglishCorpus__reading_level	458	50	58
StylePTB__relation_active_to_passive	2,434	92	209
StylePTB__relation_adjective_adverb_removal	4,284	177	401
StylePTB__relation_adjective_emphasis	625	9	62
StylePTB__relation_information_addition	1,898	80	136
StylePTB__relation_least_frequent_synonym	6,206	251	652
StylePTB__relation_most_frequent_synonym	4,132	173	415
StylePTB__relation_passive_to_active	46	0	25
StylePTB__relation_pp_back_to_front	5	1	1
StylePTB__relation_pp_front_to_back	406	0	54
StylePTB__relation_pp_removal	4,154	162	451
StylePTB__relation_substatement_removal	1,145	35	165
StylePTB__relation_to_future	6,344	256	671
StylePTB__relation_to_past	3,833	172	358
StylePTB__relation_to_present	3,615	143	419
StylePTB__relation_verb_emphasis	1,081	35	85
StylePTB__target_tense	13,601	563	77,632
StylePTB__voice	4,960	184	86,652
SynthSTEL_feature__style_relation	0	0	3,200
SynthSTEL_register__style_relation	0	0	800
ceeces__authorship_ceeces1	144	23	2,457
ceeces__authorship_ceeces2	137	0	2,487
ceeces__period	316	34	2,274
ceeces__sender_age_group	279	34	2,311
ceeces__sender_gender	317	35	2,272
core_subregister_hi	970	171	47,311
core_subregister_id	1,892	300	46,260
core_subregister_in	5,678	905	41,869
core_subregister_ip	969	162	47,321
core_subregister_ly	293	0	48,159
core_subregister_na	12,508	1,862	34,082
core_subregister_op	5,907	885	41,660
core_subregister_sp	305	53	48,094
eWAVE__variety	901	85	3,203
hate_speech__hate	8,997	1,001	946
hate_speech_and_offensive_language__offensiveness	20,685	2,298	1,800
jigsaw_toxicity_pred__identity_hate	143,591	15,954	153,190
jigsaw_toxicity_pred__insult	143,591	15,954	153,190
jigsaw_toxicity_pred__obscene	143,591	15,954	153,190
jigsaw_toxicity_pred__severe_toxic	143,591	15,954	153,190
jigsaw_toxicity_pred__threat	143,591	15,954	153,190
jigsaw_toxicity_pred__toxicity	143,591	15,954	153,190
machine_text_detection_MAGE_cmv__generator_family	13,027	1,683	19,751
machine_text_detection_MAGE_cmv__origin	18,692	4,308	11,461
machine_text_detection_MAGE_eli5__generator_family	21,397	2,686	30,832
machine_text_detection_MAGE_eli5__origin	40,914	6,182	7,819
machine_text_detection_MAGE_hswag__generator_family	21,303	2,641	16,332
machine_text_detection_MAGE_hswag__origin	27,468	6,301	6,507
machine_text_detection_MAGE_roct__generator_family	22,308	2,777	16,641
machine_text_detection_MAGE_roct__origin	28,597	6,422	6,707
machine_text_detection_MAGE_sci_gen__generator_family	15,113	1,881	15,765
machine_text_detection_MAGE_sci_gen__origin	22,297	4,734	5,728
machine_text_detection_MAGE_squad__generator_family	14,997	1,787	28,979
machine_text_detection_MAGE_squad__origin	33,336	4,623	7,804
machine_text_detection_MAGE_tldr__generator_family	17,260	2,076	13,236
machine_text_detection_MAGE_tldr__origin	22,423	4,885	5,264
machine_text_detection_MAGE_wp__generator_family	20,365	2,616	20,628
machine_text_detection_MAGE_wp__origin	29,366	6,085	8,158
machine_text_detection_MAGE_xsum__generator_family	22,673	2,841	18,311
machine_text_detection_MAGE_xsum__origin	30,392	6,500	6,933
machine_text_detection_MAGE_yelp__generator_family	13,988	1,732	47,056
machine_text_detection_MAGE_yelp__origin	42,110	4,152	16,514
masc_text_genre__genre	0	0	392
pan24_style_change_easy__authorship_relation	4,667	0	8,856
pan24_style_change_hard__authorship_relation	13,733	0	9,403
pan24_style_change_medium__authorship_relation	6,951	0	19,546
pan25_style_change_easy__authorship_relation	10,736	0	47,913
pan25_style_change_hard__authorship_relation	26,607	0	35,090
pan25_style_change_medium__authorship_relation	14,266	0	57,310
pan26_style_change_easy__authorship_relation	490,912	0	175,772
pan26_style_change_hard__authorship_relation	526,611	0	190,321
pan26_style_change_medium__authorship_relation	1,181,121	0	396,469
telegram-spam-ham__spam	17,208	1,913	1,227
twitter-airline-sentiment__sentiment	11,342	1,278	2,020
x_genre__genre	1,119	521	1,316