GottBERT: A pure German language model

GottBERT is the first German-only RoBERTa model, pre-trained on the German portion of the first released OSCAR dataset. This model aims to provide enhanced natural language processing (NLP) performance for the German language across various tasks, including Named Entity Recognition (NER), text classification, and natural language inference (NLI). GottBERT has been developed in two versions: a base model and a large model, tailored specifically for German-language tasks.

• Model Type: RoBERTa
• Language: German
• Base Model: 12 layers, 125 million parameters
• Large Model: 24 layers, 355 million parameters
• License: MIT

This was presented in GottBERT: a pure German Language Model.

Pretraining Details

• Corpus: German portion of the OSCAR dataset (Common Crawl).
• Data Size:
  • Unfiltered: 145GB (~459 million documents)
  • Filtered: 121GB (~382 million documents)
• Preprocessing: Filtering included correcting encoding errors (e.g., erroneous umlauts), removing spam and non-German documents using language detection and syntactic filtering.

Filtering Metrics

• Stopword Ratio: Detects spam and meaningless content.
• Punctuation Ratio: Detects abnormal punctuation patterns.
• Upper Token Ratio: Identifies documents with excessive uppercase tokens (often noisy content).

Training Configuration

• Framework: Fairseq
• Hardware:
  • Base Model: 256 TPUv3 pod/128 TPUv4 pod
  • Large Model: 128 TPUv4 pod
• Training Time:
  • Base Model: 1.2 days
  • Large Model: 5.7 days
• Batch Size: 8k tokens
• Learning Rate:
  • Base: Peak LR = 0.0004
  • Large: Peak LR = 0.00015
• Training Iterations: 100k steps with a 10k warm-up phase

Evaluation and Results

GottBERT was evaluated across various downstream tasks:

• NER: CoNLL 2003, GermEval 2014
• Text Classification: GermEval 2018 (coarse & fine), 10kGNAD
• NLI: German subset of XNLI

Mertics:

• NER and Text Classification: F1 Score
• NLI: Accuracy

Details:

• bold values indicate the best performing model within one architecure (base, large), undescored values the second best.

Model Architecture

• Base Model: 12 layers, 125M parameters, 52k token vocabulary.
• Large Model: 24 layers, 355M parameters, 52k token vocabulary.

Tokenizer

• Type: GPT-2 Byte-Pair Encoding (BPE)
• Vocabulary Size: 52k subword tokens
• Trained on: 40GB subsample of the unfiltered German OSCAR corpus.

Limitations

• Filtered vs Unfiltered Data: Minor improvements seen with filtered data, but not significant enough to justify filtering in every case.
• Computation Limitations: Fixed memory allocation on TPUs required processing data as a single stream, unlike GPU training which preserves document boundaries. Training was performed in 32-bit mode due to framework limitations, increasing memory usage.

Fairseq Checkpoints

Get the fairseq checkpoints here.

Citations

If you use GottBERT in your research, please cite the following paper:

@inproceedings{scheible-etal-2024-gottbert,
    title = "{G}ott{BERT}: a pure {G}erman Language Model",
    author = "Scheible, Raphael  and
      Frei, Johann  and
      Thomczyk, Fabian  and
      He, Henry  and
      Tippmann, Patric  and
      Knaus, Jochen  and
      Jaravine, Victor  and
      Kramer, Frank  and
      Boeker, Martin",
    editor = "Al-Onaizan, Yaser  and
      Bansal, Mohit  and
      Chen, Yun-Nung",
    booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.emnlp-main.1183",
    pages = "21237--21250",
}