---
license: cc-by-sa-4.0
language:
- de
pretty_name: BabyLM (German)
size_categories:
- 10M<n<100M
---

# German BabyLM dataset

This is a pre-training dataset for training developmentally plausible language models in German (also called *BabyLMs*), compiled by the Computational Linguistics Group ([CLAUSE](https://clause-bielefeld.github.io/)) at Bielefeld University.

If you are looking for ways to evaluate your German BabyLMs, we recommend [our own lexical decision dataset](https://huggingface.co/datasets/bbunzeck/lexical-decision-german), [CLAMS](https://github.com/aaronmueller/clams) for syntactic evaluation and [XCOMPS](https://www.arxiv.org/abs/2502.19737) for conceptual semantics/world knowledge.

The composition is inspired by the original, English BabyLM dataset (see [babylm.github.io](https://babylm.github.io)), more info can be found in [our paper](https://arxiv.org/abs/2503.11593), see also below for bib reference). It contains **16.5M words** from a variety of sources:


| Dataset | Description | \# Words |
|---|---|---:|
| CHILDES | Child-directed speech | 3,626,301 |
| CHILDES | Child speech | 1,511,144 |
| OpenSubtitles | Movie subtitles | 1,543,094 |
| CallHome  | Phone conversations | 176,313 |
| Klexikon | Children's online encyclopedia | 1,384,891 |
| MiniKlexikon | Simplified online encyclopedia | 272,886 |
| Wikibooks Wikijunior | Educational books | 226,773 |
| Fluter | German youth magazine | 2,862,278 |
| Project Gutenberg | Literature (children's and young adult) | 2,476,133 |
| Dreambank  | Dream reports | 939,197 |
| Leipzig corpus news texts | Short news texts | 1,541,803 |
| **Total** | | **16,560,813** |

The CHILDES data were taken from [the CHILDES database for German](https://childes.talkbank.org/access/German/).

OpenSubtitles data was sourced from [OPUS](https://opus.nlpl.eu/OpenSubtitles/corpus/version/OpenSubtitles).

CallHome data was taken from [TalkBank](https://ca.talkbank.org/access/CallHome/deu.html).

We collected the [Klexikon](https://huggingface.co/datasets/bbunzeck/klexikon), [MiniKlexikon](https://huggingface.co/datasets/bbunzeck/mini-klexikon), [Wikibooks](https://huggingface.co/datasets/bbunzeck/wikibooks-wikijunior) and [Fluter](https://huggingface.co/datasets/bbunzeck/fluter) data ourselves.

Book data was downloaded from [Project Gutenberg](https://www.gutenberg.org/).

Dreambank data was extracted from [this dataset](https://huggingface.co/datasets/DReAMy-lib/DreamBank-dreams).

The news texts were downloaded from the [Leipzig corpora](https://wortschatz.uni-leipzig.de/de/download/German).

If you use our BabyLM dataset, please cite the following preprint:

```
@misc{bunzeck2025constructiondistributionsshapeformal,
      title={Do Construction Distributions Shape Formal Language Learning In German BabyLMs?}, 
      author={Bastian Bunzeck and Daniel Duran and Sina Zarrieß},
      year={2025},
      eprint={2503.11593},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.11593}, 
}
```