---
language:
- da
license: cc-by-4.0
task_categories:
- text-generation
task_ids:
- language-modeling
tags:
- danish
- statistics
- data-to-text
- instruction-tuning
- Danmarks Statistik
pretty_name: DST Table Prompts
size_categories:
- 1K<n<10K
---

# DST Table Prompts

A Danish instruction-tuning dataset of `(prompt, article)` pairs derived from
[Danmarks Statistik](https://www.dst.dk) (Statistics Denmark) publications.

Each example pairs a natural-language user request — embedding the actual
markdown table — with the real statistician-written article as the target
response. The prompts are LLM-generated and vary in style, tone, and table
placement; the table data and article text come directly from the source
dataset.

## Dataset description

Danmarks Statistik publishes statistical articles alongside the raw data tables
they are based on. This dataset turns each article into a data-to-text training
example:

- **prompt**: a natural Danish user request asking an AI to write an article
  from the table data. The request embeds the real markdown table(s) from the
  article.
- **target**: the original article written by Danmarks Statistik statisticians.

The prompts were generated with `gpt-4.1` using 12 cycling style seeds (varying
tone and table placement) and personas sampled from
[nvidia/Nemotron-Personas-USA](https://huggingface.co/datasets/nvidia/Nemotron-Personas-USA).
The LLM never sees or modifies the table data — it generates framing text
around a `[table]` placeholder that is replaced with the actual markdown in
post-processing.

## Source data

Built from [`oliverkinch/dst-table-prompts`](https://huggingface.co/datasets/oliverkinch/dst-table-prompts),
which contains DST statistical articles paired with their inline markdown tables
scraped from [dst.dk/nyt](https://www.dst.dk/nyt).

## Dataset structure

| Split | Rows |
|-------|------|
| train | 3 043 |
| test  | 339 |

### Fields

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Unique row identifier (`dst_tables-NNNNNN`) |
| `prompt` | string | Danish user request with the markdown table(s) embedded |
| `target` | string | Original article text from Danmarks Statistik |
| `sources` | list | Provenance: source dataset, split, row id, title, URL |
| `meta` | dict | Metadata: title, URL, char counts, style seed index, source row index |
| `checks` | dict | Quality checks run at verification time |

### Example

```python
from datasets import load_dataset
ds = load_dataset("oliverkinch/dst-table-prompts")
print(ds["train"][0]["prompt"][:300])
print("---")
print(ds["train"][0]["target"][:300])
```

## Generation

Built with [dfm-sdg](https://github.com/danish-foundation-models/dfm-sdg)
(`dst_tables` pack). Run ID: `2026-05-04T131736+0000-fa019553`.

## Licence

The articles and tables originate from Danmarks Statistik and are published
under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). The generated
prompts are derived works and inherit the same licence.