NbAiLab/nb-asr-supermorphed-nno

This dataset teaches direct conversion from ordinary Norwegian Nynorsk text to AltMorph bracket encoding. Every clean row contains an original source and one validated target. Allowed alternatives inside a block are alphabetized and use the form [alternativ1|alternativ2].

Splits

Split	Rows	Source words
train	482,592	9,457,601
validation	1,000	18,257
test	1,000	18,701

Validation and test contain exactly 1,000 rows each. They were selected only after cleaning with a fixed random document-level hash. No other row from a held-out upstream document appears in train. Test contents were not manually inspected after assignment.

Provenance and licenses

Each row retains source_dataset, immutable source_revision, source_row_id, source_document_id, source_type, and license. The corpus contains material under several compatible open licenses, so the Hub-level license is marked other; the authoritative license for reuse is the row-level license value. Consult the upstream dataset cards as well.

Construction

The build used only upstream train material. Text was NFC- and whitespace-normalized and split into rows of at most 200 whitespace words during collection. Collected rows remain available for later rounds, while this release selects at most 10 million source words and excludes rows over 120 words. Only exact normalized duplicates were removed, with lineage retained in the private build audit. Målfrid was excluded by provenance after source-quality analysis and contributes no release rows.

Every deduplicated row was scored by the pinned NbAiLab/nb-nordic-lid models nb-nordic-lid.159.ftz and nb-nordic-lid.ftz at revision 12401f3e0d4e73adf86a089e828d82d37242ccab (OpenRAIL). LID was used only as a high-confidence foreign-text veto; it never assigned Bokmål versus Nynorsk. Either Norwegian label passed. Broad non-Norwegian predictions required at least 12 words and confidence 0.99. Danish, Swedish, Faroese, and Icelandic required at least 40 words and matching 0.99 predictions from both models. Raw LID labels and scores are retained in the private build audit and are not published as dataset columns.

A pinned multilingual E5 encoder assigned deterministic cosine-LSH clusters. Sampling used source type, length, and semantic cluster jointly, with a document cap outside parliamentary transcripts and a square-root length cost that mildly favors shorter rows without imposing a percentile quota.

Two candidate generators were compared in a teacher/student workflow. The rule-based teacher candidate used AltMorph revision 895347c82c07f6e4f7e2d3afec838291838c4c87, HumIT Humit-Oslo/humit-tagger-large at revision 942f2901ec3271c9f627eb6a78b07ade87cf8126, Ordbank with mandatory complete inflection search, and North-T5 north/t5_base_NCC at revision 23788220cee2fb254bb0aa175074a5dd03f8d763.

The launch-time Git worktree dirty flag was true; the authoritative implementation fingerprint was 1b32b0f68f2bb94e513feefba439ae075c701875e12e9989214aeab1256b0328 and the run-contract SHA-256 was b55bec8a2e0f7b50fbb390b0e1836b11c60a5c91ae33eb8c68fe4496202cc71e. Generation used 5 shards, batches of 100 rows, HumIT/North-T5 batches of 16/16, and 4 local cache-lookup threads per worker. Corpus generation was cache-only and could not contact Ordbank.

The model candidate used google/t5gemma-2-1b-1b at revision dd0a2683227859151b1730ca3a63087df5b5f39b. A separate model for nno was fine-tuned for exactly 1 epoch on the complete strict AltMorph teacher corpus local-supermorph-teacher-nno, pinned by manifest SHA-256 4ce59d483bfeb99f052179cfceb5dc814e0f8e9623a3f47de966fa4f9893987b. Its deterministic teacher_eval view overlaps training and is only a consistency diagnostic, not held-out validation. Run t5gemma2-1b-supermorph-nno-1ep-l40s-micro4-math-v1 used the final checkpoint at step 3906 (final checkpoint after exactly one language-specific epoch) and reached exact match 0.961000. Greedy inference used batch size 8, with source/target limits of 512/512 tokens. The resolved training configuration SHA-256 was 5ffa2da9d33a1045cd99f36f62830c51a486e0172d545dd7c9fbcafe6e7956e3.

Valid exact agreements were accepted automatically. Their rate was 96.90%; blind-review packet creation required at least 90.00%. The 15,486 disagreements were shown independently to three blind reviewers. Candidate provenance was hidden and A/B order was randomized separately for every reviewer. Reviewers could select only A, B, or None and could not edit text. A disagreement was retained only when all three selected the same underlying structurally valid candidate.

The audit configuration preserves rejected rows, both candidates, decoded votes, and rejection status. It also preserves accepted companions excluded solely to prevent held-out-document leakage. These rows can be reconsidered by future models; they were not silently deleted.

Cleaning summary

• automatic valid agreements: 484,384
• accepted by three-reviewer consensus: 452
• rejected or unresolved before split isolation: 15,034
• held-out document companions moved to audit: 244

Build scripts and the complete reproducibility contract are maintained in the AltMorph repository under dataset_tools/supermorphed/.