NbAiLab/nb-asr-morfologisk-stil-nob

Document-conditioned Bokmål morphological style

This dataset trains a model to rewrite one sentence in the morphological style described by 5–20 observed word forms from the same source document. Every listed form has verified alternatives in AltMorph. Evidence is deduplicated, comes from another sentence, and is disjoint from the target's alternative families.

Task format

The prompt is ready for sequence-to-sequence training:

skriv_i_samme_morfologiske_stil:
morfologiske_stilvalg: hytta | hosta | ...
setning_med_alternativer: [Jenta|Jenten] [kasta|kastet] ballen.

The second input mode uses setning: followed by a deliberately scrambled surface sentence. The target is always only the completed sentence; no hidden reasoning or structured intermediate target is required.

The causal config contains 28,791 matched pairs. Each pair holds the bracketed sentence fixed while two different reference documents provide contrasting style lists and license two different completed sentences. Pair accuracy therefore cannot be solved by a single global majority form.

Splits

split	rows
train	327,912
validation	34,032
test	34,762

Splits are assigned by a fixed hash over near-duplicate document clusters. No source document or detected near-duplicate cluster crosses a split. Test was assigned without target inspection and is excluded from the human-inspection packet.

Provenance and licenses

The source is NbAiLab/nb-asr-supermorphed-nob at revision 346754624582a2858864a18c685aa365d338504e. Every row retains the upstream dataset, revision, row/document identifiers, source type, and row-level license. The combined corpus uses multiple compatible open licenses, so the Hub-level license is other; consumers must use each row's license field and consult the upstream dataset cards.

Construction and audit

The release contains 396,706 rows derived from 84,101 distinct target sentences. Alternative blocks were aligned exactly to the human-written source. Style forms were kept only when every candidate was a single lexical form. Literal Markdown, mathematical, and editorial brackets were preserved rather than treated as AltMorph syntax.

Each row includes the alternatives and evidence sentence for every style form, plus target-family metadata and near-duplicate cluster ID. manifest.json, audit.json, source_profile.json, and human_inspection.json provide the reproducibility and inspection records.

Limitations

The style list is evidence, not a guaranteed globally coherent linguistic label. Source documents can contain quotations, multiple authors, or genuine variation. Evaluation should therefore report bracketed and scrambled modes separately, compare unchanged and rule-based baselines, and include a matched contrasting-style-list causal test.