Add dataset card

README.md

@@ -1,51 +1,80 @@
 ---
-dataset_info:
-  features:
-  - name: id
-    dtype: string
-  - name: meta
-    struct:
-    - name: content_type
-      dtype: string
-    - name: passage_idx
-      dtype: int64
-    - name: source_dataset
-      dtype: string
-    - name: source_id
-      dtype: string
-    - name: source_row_index
-      dtype: int64
-    - name: source_split
-      dtype: string
-    - name: target_chars
-      dtype: int64
-    - name: title
-      dtype: string
-  - name: prompt
-    dtype: string
-  - name: sources
-    list:
-    - name: dataset
-      dtype: string
-    - name: row_id
-      dtype: string
-    - name: split
-      dtype: string
-    - name: title
-      dtype: string
-    - name: url
-      dtype: string
-  - name: target
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 9289968
-    num_examples: 7154
-  download_size: 6070062
-  dataset_size: 9289968
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
+language:
+- da
+license: cc-by-4.0
+task_categories:
+- question-answering
+- text-generation
+tags:
+- danish
+- statistics
+- synthetic
+- instruction-tuning
+- backtranslation
+pretty_name: Danmarks Statistik BT
+size_categories:
+- 1K<n<10K
+source_datasets:
+- oliverkinch/danmarks-statistik
 ---
+
+# Danmarks Statistik BT
+
+Synthetic Danish instruction-tuning dataset built from [Danmarks Statistik](https://www.dst.dk) publications using backtranslation. Each row pairs a short, natural Danish chatbot input (`prompt`) with a prose passage from a DST publication as the grounding answer (`target`).
+
+## Dataset construction
+
+Passages are extracted from the source dataset [`oliverkinch/danmarks-statistik`](https://huggingface.co/datasets/oliverkinch/danmarks-statistik), which covers four content types published by Danmarks Statistik:
+
+| Content type | Description | Rows |
+|---|---|---|
+| `bag` | Background material and in-depth analysis | 3,190 |
+| `nyt` | Short statistical news releases | 1,926 |
+| `pub` | Statistical publication summaries | 1,416 |
+| `analyse` | In-depth statistical analyses | 622 |
+
+For each passage an LLM writes a short, persona-conditioned Danish chatbot input that a person might send about the topic — without referencing specific numbers or the source document. Personas are sampled from [`nvidia/Nemotron-Personas-USA`](https://huggingface.co/datasets/nvidia/Nemotron-Personas-USA) to produce varied registers (researchers, students, journalists, retirees, etc.).
+
+Generated prompts are filtered to ensure they:
+- end with terminal punctuation
+- do not reference "the above text" or similar
+- are in Danish (not English)
+
+## Statistics
+
+| | |
+|---|---|
+| Split | train |
+| Rows | 7,154 |
+| Prompt length (chars) | min 42 · mean 129 · max 367 |
+| Target length (chars) | min 200 · mean 686 · max 2,499 |
+| Generation failures | 8 (0.1%) |
+
+## Schema
+
+| Field | Type | Description |
+|---|---|---|
+| `id` | string | Unique row identifier, e.g. `dst-000013-p00` |
+| `prompt` | string | Short Danish chatbot input about the passage topic |
+| `target` | string | Source passage from a DST publication |
+| `sources` | list | Provenance: source dataset, split, row id, title, URL |
+| `meta` | dict | Content type, passage index, character count, source coordinates |
+
+## Example
+
+```python
+from datasets import load_dataset
+ds = load_dataset("oliverkinch/danmarks-statistik-bt", split="train")
+print(ds[0]["prompt"])
+# "Har du en fornemmelse af, hvorfor eksporten er faldet igen ..."
+print(ds[0]["target"])
+# "I februar faldt den samlede eksport af varer og tjenester med 3,7 pct. ..."
+```
+
+## Intended use
+
+Fine-tuning and evaluating Danish language models on instruction-following and retrieval-augmented generation tasks grounded in official Danish statistics.
+
+## Source data
+
+Danmarks Statistik publications are released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). See the [source dataset](https://huggingface.co/datasets/oliverkinch/danmarks-statistik) for full provenance.