README.md

metadata

language:
  - nb
license: cc-by-sa-3.0
task_categories:
  - text-generation
pretty_name: NPSC Ortho Morphcoded Clean
tags:
  - norwegian
  - morphology
  - text2text-generation
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: eval
        path: data/eval-*
      - split: eval_clean
        path: data/eval_clean-*
      - split: test
        path: data/test-*
      - split: test_clean
        path: data/test_clean-*

NPSC Ortho Morphcoded Clean

This is a consensus-filtered derivative of NbAiLab/NPSC_orto_morphcoded at revision 'c0a5864ffde32ca05a683652b54282ee785ca16d'. It retains the original id, source, and target schema and the original source text.

Cleaning method

A T5Gemma 2 1B model fine-tuned on the source dataset generated one prediction for every row. Exact model/target agreements were retained. Every disagreement was shown to two isolated language-model reviewers with the source plus randomized candidates A and B; neither reviewer saw candidate provenance or the other review. Each reviewer could select only A, B, or None.

A disagreement was retained only when both reviewers independently selected the same non-None candidate after candidate randomization was decoded. Every other disagreement was dropped. A unanimous candidate was also dropped if it failed AltMorph syntax or exact source-recoverability checks. Selected targets were canonicalized to case-insensitive alphabetical alternative order.

The automated reviewers are a quality-control heuristic, not human linguistic annotation. Agreement between the generator, model, or reviewers does not prove that every retained alternative is authoritative or context-preserving.

Source split	Input	Model/target agreement	GT selected	Model selected	Dropped
train	62,541	61,346	566	482	147
validation	1,000	924	45	22	9
test	1,000	934	38	21	7

Published splits

train is the cleaned original training split. eval and test preserve all original evaluation rows and labels for direct comparability. eval_clean and test_clean contain only rows retained by the consensus-cleaning procedure, with the selected target.

Split	Rows	Source differs from target
train	62,394	35,830
eval	1,000	581
eval_clean	991	582
test	1,000	563
test_clean	993	561

Split overlap

Split IDs are disjoint, but the upstream random split contains repeated rows. The following counts are measured against train after cleaning. Exact repeated source/target pairs can make evaluation optimistic and must not be interpreted as independent generalization.

Split	ID overlap rows	Exact source/target overlap rows	Unique overlapping sources
eval	0	105	84
eval_clean	0	105	84
test	0	99	76
test_clean	0	99	76

Fields

Field	Meaning
id	Original NPSC sentence ID as a string.
source	Norwegian parliamentary transcript text.
target	Source-preserving AltMorph encoding with alphabetically ordered alternatives.

Provenance

• Generated: 2026-08-29T00:00:23.289993+00:00
• Source dataset: 'NbAiLab/NPSC_orto_morphcoded'
• Source revision: 'c0a5864ffde32ca05a683652b54282ee785ca16d'
• Cleaning model: google/t5gemma-2-1b-1b, fine-tuned run t5gemma2-1b-altmorph-v2-alpha
• Reviewers: two independently isolated automated language-model agents
• Decision choices: A, B, or None
• Retention rule for disagreements: unanimous A or unanimous B only

The source dataset card documents NPSC and AltMorph generation in detail, including HumIT, Ordbank, North-T5, source licensing, and attribution. This derivative uses CC BY-SA 3.0 conservatively and preserves those attribution requirements.

Please credit:

• the National Library of Norway and NPSC dataset maintainers;
• the University of Oslo HumIT team and Humit-Oslo/humit-tagger-large;
• the University of Bergen Ordbank service;
• the North-T5 authors and north/t5_base_NCC;
• the AltMorph authors; and
• Google and the T5Gemma authors for the cleaning model.

The upstream parliamentary transcriptions are described as CC0, while the Hugging Face curation and this derivative use CC BY-SA 3.0.

Limitations

The dataset remains automatically generated silver data. In-sample predictions on the training split can reflect memorization, and shared model/target errors survive automatic agreement filtering. Agent judgments can also be correlated or linguistically wrong. Use eval/test for comparison with the original silver labels and eval_clean/test_clean for the stricter retained subset. Fresh human or linguist evaluation is recommended for consequential use.