peterkirby/mymirmdatasets2

MUSE Style and Complexity Datasets

CEFR-SP

cefr_sp contains the publicly redistributable SCoRE and Wiki-Auto portions of the CEFR-based Sentence Profile corpus. The gated Newsela-Auto portion is not included. The source's train and test splits are retained, and dev is exposed as validation.

The source sentences were rated independently by two English-education professionals. Both ratings are retained as rating_1 and rating_2. For MUSE, the six CEFR levels are mapped to four ordered bands:

primary_id_int	CEFR band
0	A1-A2
1	B1
2	B2
3	C1-C2

The primary label is the lower of the two mapped ratings. Thus A2/B1 maps to A1-A2, B1/B2 maps to B1, and B2/C1 maps to B2. This consistently treats an annotator disagreement as evidence for the lower complexity band.

Columns

• text: source sentence, unchanged except for removal of the file line ending
• primary_id_int: four-band complexity label
• secondary_category_int: always -1 (shared-content group unknown)
• source_record_id: source corpus, original split, and zero-based source row
• rating_1: first annotator's CEFR rating
• rating_2: second annotator's CEFR rating

One repeated Wiki-Auto training sentence with the same two ratings was removed, keeping its first source occurrence. No balancing or synthetic adjudication was applied.

Split	Rows	A1-A2	B1	B2	C1-C2
train	7,177	2,058	2,712	1,980	427
validation	1,366	516	457	274	119
test	1,460	525	534	280	121
Total	10,003	3,099	3,703	2,534	667

Source and licensing

CEFR-SP is distributed by Yuki Arase, Satoru Uchida, and Tomoyuki Kajiwara in the CEFR-SP repository. This release is pinned to commit 2ba4a5005fbff68ab0f863891cf977c50fdcda29.

There is no single content license for this combined configuration:

• Wiki-Auto rows retain the source's CC BY-SA 3.0 terms.
• SCoRE rows retain the source's CC BY-NC-SA 4.0 terms.

The source component is encoded in every source_record_id. Users are responsible for complying with the applicable source-specific license, including attribution, share-alike, and non-commercial restrictions.

Please cite:

@inproceedings{arase-etal-2022-cefr,
  title = {{CEFR}-Based Sentence Difficulty Annotation and Assessment},
  author = {Arase, Yuki and Uchida, Satoru and Kajiwara, Tomoyuki},
  booktitle = {Proceedings of EMNLP 2022},
  year = {2022},
  pages = {6206--6219},
  doi = {10.18653/v1/2022.emnlp-main.416}
}

CLMET 3.1

The Corpus of Late Modern English Texts 3.1 contains 333 complete works by British and Irish authors, covering 1710-1920. It supplies explicit genre and historical-period metadata. The two configurations use the same works and split assignment but expose different supervised labels.

One source work is one row. The metadata header, paragraph tags, and page markers are removed. The remaining CLMET plain text is preserved, including its historical spelling, punctuation, whitespace tokenization, and line breaks. secondary_category_int is -1 because a work occurs only once within each configuration.

The source has no official train/test division. These configurations use a frozen, approximately 90/10 train-validation split. All works with the same raw author value stay together, and the identical assignment is used in both configurations. This prevents author leakage while retaining at least two validation works in every supervised class.

clmet_genre

This configuration uses the five major genres named by the corpus creators. The 37 unclassified Other works are excluded rather than being treated as a coherent sixth genre. The source's finer subgenre remains metadata only: it contains sparse and hybrid values that are not suitable as a standalone MUSE task.

primary_id_int	Major genre	Train	Validation
0	Narrative fiction	114	13
1	Narrative non-fiction	24	3
2	Drama	67	7
3	Letters (raw value LET)	11	2
4	Treatise	50	5
	Total	266	30

The work counts are uneven, particularly for letters. A class-aware sampler is recommended for contrastive training; the release does not duplicate or downsample works.

clmet_period

This configuration includes all 333 works and uses the official three 70-year periods.

primary_id_int	Historical period	Train	Validation
0	1710-1780	79	9
1	1780-1850	89	10
2	1850-1920	132	14
	Total	300	33

The 37 Other-genre works are retained here because their historical-period labels are fully defined: 34 are in train and 3 in validation.

CLMET columns

• text: complete work after removal of CLMET structural markup
• primary_id_int: major genre or historical period, depending on configuration
• secondary_category_int: always -1
• source_record_id: clmet3.1:<source id>
• author, title, year: source bibliographic metadata
• period, quarter_cent, decade: source historical metadata
• genre, subgenre: raw source genre metadata

CLMET source and licensing

CLMET 3.1 was created by Hendrik De Smet, Susanne Flach, Hans-Jürgen Diller, and Jukka Tyrkkö. Data come from the canonical CLMET 3.1 release. The archive is pinned by its published MD5 and SHA-1 checksums and an additional SHA-256 recorded by the builder.

CLMET 3.1 is licensed under CC BY-NC-SA 4.0. Its attribution, non-commercial, and share-alike requirements apply to both CLMET configurations.

Please cite the corpus creators and:

Diller, Hans-Jürgen, Hendrik De Smet, and Jukka Tyrkkö. 2011. A European database of descriptors of English electronic texts. The European English Messenger 19: 21-35.

SynthSTEL-Luna

Forty feature-specific synthstel_luna__* configurations expose 4,000 fresh English contrast pairs as 8,000 individual sentence rows. Each config contains 180 training sentences (90 aligned pairs) and 20 validation sentences (10 pairs). The source's train/test assignments are unchanged; only the uploaded split name test becomes validation. Original pair exports and sidecars are retained in supporting/synthstel_luna/exports/ and are not extra training configurations.

primary_id_int=0 means the supplied negative side and 1 the supplied positive side of that particular contrast. These are feature-side classes, not universal sentiment or quality labels, and the numbers alone do not define ordinal loss. The readable primary_label records the original condition. The two members of each pair share secondary_category_int, numbered 0–99 by sorted original pair ID within that config. IDs are scoped to the configuration across both splits. No pair or content family crosses splits. All 40 raw contrasts are retained. The inventory's cleaned labels merge some perspective contrasts; the final pair exports instead retain raw names for four perspective contrasts. The published feature_clean preserves each actual exported value, and mappings retain inventory_feature_clean separately. Neither replaces the raw feature as the identity of a training config.

For contrastive training, sample within a feature config: same-side sentences from different pairs provide same-feature examples across topics; the opposite side of an aligned pair provides a content-controlled contrast. Do not pool primary IDs across configs. Preserve the more/less member roles when constructing STEB-style order-alignment trials; these rows are a training representation, not new official STEB evaluation data. style_group identifies the eight register contrasts and 32 linguistic-feature contrasts used in STEB's grouping.

Rows retain pair/content-family IDs, member role, raw and cleaned feature names, original split, and complete generation/review provenance. The source package records generation by five reused gpt-5.6-luna workers at Medium reasoning and rotated agent review. Fast/priority was unexposed. No human validation is claimed. Actual provenance includes fresh in-run topics; C4 topic extraction is not independently established by those records. This is a methodological adaptation of StyleDistance, not an exact replication or a claim of human-rated correctness.

The retained validation report describes character 4–6-gram cosine screening at 0.97 against pinned original SynthSTEL exports; this heuristic is not exhaustive semantic duplicate detection. This publication independently rechecks casefolded, whitespace-collapsed exact matches against those originals. It does not establish comprehensive clearance against every STEB dataset. Full reports, reference pins, generation settings and review limitations accompany the supporting artifacts.

The new text is generated data; this release does not assert a blanket license for all configurations. Existing source-dependent licenses remain applicable. Original SynthSTEL is an evaluation/reference source and is not redistributed here. Cite Patel et al., StyleDistance: Stronger Content-Independent Style Embeddings with Synthetic Parallel Examples, NAACL 2025: https://aclanthology.org/2025.naacl-long.436/

Config	Feature	ID 0	ID 1	Group
synthstel_luna__complex_sentence_structure	Complex Sentence Structure	Simple	Complex	register
synthstel_luna__fluency_in_sentence_construction	Fluency in Sentence Construction	Disfluent sentence	Fluent sentence	feature
synthstel_luna__frequent_usage_of_common_verbs	Frequent Usage of Common Verbs	Less frequent common verbs	With common verbs	feature
synthstel_luna__frequent_usage_of_determiners	Frequent Usage of Determiners	Less frequent determiners	With determiners	feature
synthstel_luna__frequent_usage_of_function_words	Frequent Usage of Function Words	Less frequent function words	With function words	feature
synthstel_luna__frequent_usage_of_punctuation	Frequent Usage of Punctuation	Less frequent punctuation	With frequent punctuation	feature
synthstel_luna__incorporation_of_humor	Incorporation of Humor	Without humor	With humor	register
synthstel_luna__positive_sentiment_expression	Positive Sentiment Expression	Negative	Positive	register
synthstel_luna__presence_of_misspelled_words	Presence of Misspelled Words	Normal sentence	Sentence with a few misspelled words	feature
synthstel_luna__usage_of_active_voice	Usage of Active Voice	Passive	Active	feature
synthstel_luna__usage_of_articles	Usage of Articles	Less frequent articles	With articles	feature
synthstel_luna__usage_of_certain_tone	Usage of Certain Tone	Uncertain	Certain	register
synthstel_luna__usage_of_conjunctions	Usage of Conjunctions	Less frequent conjunctions	With conjunctions	feature
synthstel_luna__usage_of_contractions	Usage of Contractions	Without contractions	With contractions	feature
synthstel_luna__usage_of_emojis	Usage of Emojis	No emojis	With emojis	feature
synthstel_luna__usage_of_formal_tone	Usage of Formal Tone	Informal	Formal	register
synthstel_luna__usage_of_long_words	Usage of Long Words	Short average word length	Long average word length	feature
synthstel_luna__usage_of_metaphors	Usage of Metaphors	Without metaphor	With metaphor	feature
synthstel_luna__usage_of_nominalizations	Usage of Nominalizations	Without nominalizations	With nominalizations	feature
synthstel_luna__usage_of_numerical_digits	Usage of Numerical Digits	Less frequent digits	With digits	feature
synthstel_luna__usage_of_numerical_substitution	Usage of Numerical Substitution	Without number substitution	With number substitution	feature
synthstel_luna__usage_of_offensive_language	Usage of Offensive Language	Non-offensive	Offensive	register
synthstel_luna__usage_of_only_lowercase_letters	Usage of Only Lowercase Letters	Proper capitalization	All lower case	feature
synthstel_luna__usage_of_only_uppercase_letters	Usage of Only Uppercase Letters	Proper capitalization	All upper case	feature
synthstel_luna__usage_of_personal_pronouns	Usage of Personal Pronouns	Less frequent pronouns	With personal pronouns	feature
synthstel_luna__usage_of_polite_tone	Usage of Polite Tone	Impolite	Polite	register
synthstel_luna__usage_of_prepositions	Usage of Prepositions	Less frequent prepositions	With prepositions	feature
synthstel_luna__usage_of_present_focused_vs_future_focused	Usage of Present-focused vs. Future-focused	Future-focused	Present-focused	feature
synthstel_luna__usage_of_present_focused_vs_past_focused	Usage of Present-focused vs. Past-focused	Past-focused	Present-focused	feature
synthstel_luna__usage_of_pronouns	Usage of Pronouns	Less frequent pronouns	With pronouns	feature
synthstel_luna__usage_of_sarcasm	Usage of Sarcasm	Without sarcasm	With sarcasm	register
synthstel_luna__usage_of_self_focused_language_vs_audience_focused	Usage of Self-Focused Language vs. Audience-focused	Audience-focused	Self-focused	feature
synthstel_luna__usage_of_self_focused_language_vs_inclusive_focused	Usage of Self-Focused Language vs. Inclusive-focused	Inclusive-focused	Self-focused	feature
synthstel_luna__usage_of_self_focused_language_vs_you_focused	Usage of Self-Focused Language vs. You-Focused	You-focused	Self-focused	feature
synthstel_luna__usage_of_self_focused_perspective_vs_third_person_singular	Usage of Self-Focused Perspective vs. Third-person Singular	Third-person singular	Self-focused	feature
synthstel_luna__usage_of_text_emojis	Usage of Text Emojis	No emojis	Text emojis	feature
synthstel_luna__usage_of_uppercase_letters	Usage of Uppercase Letters	Without uppercase letters	With uppercase letters	feature
synthstel_luna__usage_of_words_indicating_affective_vs_cognitive_processes	Usage of Words Indicating Affective vs. Cognitive Processes	Cognitive processes	Affective processes	feature
synthstel_luna__usage_of_words_indicating_affective_vs_perceptual_processes	Usage of Words Indicating Affective vs. Perceptual Processes	Perceptual processes	Affective processes	feature
synthstel_luna__usage_of_words_indicating_cognitive_vs_perceptual_processes	Usage of Words Indicating Cognitive vs. Perceptual Processes	Perceptual processes	Cognitive processes	feature

Loading and reproduction

Every config has text, primary_id_int, secondary_category_int, and source_record_id; extra fields preserve source-specific metadata. Unknown shared content uses -1, which must never count as one shared group. CEFR-SP remains the default config. CLMET's genre/period views overlap in their source works, so summing their rows does not count unique works. Sample spans from CLMET's complete long works rather than repeatedly truncating their openings.

from datasets import load_dataset
repo = "peterkirby/mymirmdatasets2"
genre = load_dataset(repo, "clmet_genre")
period = load_dataset(repo, "clmet_period")
contrast = load_dataset(repo, "synthstel_luna__usage_of_formal_tone")

The release has 43 configs and 18,632 configuration rows: 10,003 CEFR-SP, 629 across the two overlapping CLMET views, and 8,000 SynthSTEL-Luna sentences. Use reproduction/README.md for publisher commands and source-package layout. release.json and validation.json record source pins, checksums and local acceptance. The publication receipt returned by the publisher records the exact Hub revision and remote verification. No model training is performed.