oliverkinch/dynaword-no-bt

Dataset Card for oliverkinch/dynaword-no-bt

Dataset Summary

dynaword-no-bt is a Norwegian instruction-tuning dataset generated with backtranslation from selected subsets of danish-foundation-models/norwegian-dynaword.

Each row contains:

• —prompt: a synthetic Norwegian user request suitable for instruction fine-tuning
• —target: the source text passage that the prompt is intended to elicit
• —meta and sources: provenance metadata (source subset, source row id, split, source type)

The dataset was built with the backtranslation_passages_dynaword_nb pipeline in dfm-sdg and includes source-aware prompting, prompt form rotation, and passage-level filtering.

Dataset Description

• —Number of rows: 6,013 (train)
• —Avg prompt length: 155.1 chars
• —Avg target length: 1,940.7 chars
• —Source corpus: danish-foundation-models/norwegian-dynaword

Source distribution

• —government-nob: 1,240
• —maalfrid: 1,203
• —wikipedia-nob: 991
• —wikipedia-nno: 967
• —government-nno: 876
• —public-reports: 736

Data Fields

• —id (string): synthetic row id
• —prompt (string): generated Norwegian instruction-like user prompt
• —target (string): source text/passage
• —meta (struct): pipeline metadata including
• —source_name, source_type, source_id, source_record_index, passage_idx, target_chars
• —sources (list[struct]): provenance metadata (dataset, config_name, split, row_id)

Creation Process

Rows were generated with source-specific prompting styles:

• —government: maalfrid, government-nob, government-nno, public-reports
• —encyclopedic: wikipedia-nob, wikipedia-nno

Pipeline highlights:

• —hybrid/passages chunking with per-source controls
• —wikitext markup stripping for Wikipedia sources
• —filtering for OCR noise, repetitive segments, and sensitive content
• —prompt form rotation (question, direct command, conversational, task context)
• —prompt post-processing and validation (Norwegian language checks, prompt-shape checks, leakage controls, enumeration rejection)

Build verification checks passed for all uploaded rows:

• —prompt_present
• —target_present
• —target_min_chars