transformations-danish-danish

Chat-template-ready synthetic instruction-transformation examples generated with google/gemma-4-31B-it. Each example transforms Danish source text into a Danish response.

Contents

• Format: gzip-compressed JSON Lines under data/*.jsonl.gz
• Schema: {"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}
• Files: 250
• Rows: 208,117
• Language pair: da_da

Actual Sources

The exported rows contain only chat messages; aggregate provenance below was reconstructed by exact matching against accepted generation metadata.

Source cluster	Rows
danish-foundation-models/danish-dynaword	146,043
danish-foundation-models/laerebogen	33,837
synquid/wiki-instruct-da	8,740
oliverkinch/instruct-bt	5,033
oliverkinch/tidsskrift-dk-bt	4,843
LexDK converted sources	4,715
oliverkinch/danmarks-statistik-bt	4,604
oliverkinch/doab-da-bt	302

Task Families

Task family	Rows
numbered_fact_extraction	49,611
non_copy_rewrite	47,023
exact_sentence_summary	45,270
child_friendly_simplification	44,156
past_tense_rewrite	22,057

Export Checks

Check	Rows
Exported rows	208,117
Matched to generation metadata	208,117
Unmatched exported rows	0
Accepted regeneration rows in export	0

Example

{
  "messages": [
    {
      "role": "user",
      "content": "Instruktion: Forklar teksten med enkle hverdagsord. Kildesprog: dansk. Svarsprog: dansk. Outputkrav: Returnér kun den forenklede tekst. Skriv ikke formuleringer som 'Her er' eller 'Det betyder'. Passage: På fransk er 'app' (forkortelse for 'application') et feminint substantiv, så det hedder 'une application'. Derfor bruges den feminine form: 'une nouvelle application'. Kønnet er ofte svært at gætte ud fra lyd eller betydning – det skal ofte bare læres."
    },
    {
      "role": "assistant",
      "content": "På fransk er ordet for \"app\" et hunkønsord. Derfor siger man \"une nouvelle application\". Det kan være svært at gætte, om et ord er hankøn eller hunkøn, så man er ofte nødt til at lære det udenad."
    }
  ]
}

Filtering

Rows were generated and judged with google/gemma-4-31B-it. Rejected or unjudged rows were excluded; accepted regeneration rows replace their original generated rows.

Exact byte-for-byte recreation is not expected because teacher sampling, serving order, and judge decisions can vary. Seed selection, task order, and prompt templates are deterministic for a fixed seed.

Machine-readable details are in generation_config.json, accepted_selection_summary.json, and audit_summary.json.

Recreate

Start an OpenAI-compatible teacher endpoint, then run a small smoke sample:

python recreate_dataset.py \
  --base-url http://127.0.0.1:8100/v1 \
  --model posttrain-gemma-teacher \
  --rows 1000 \
  --output generated/train.jsonl.gz

By default, the script uses seeds/source_texts.jsonl.gz and writes only judge-accepted rows.