Edit model card

GPT-Neo 125M pre-trained on cleaned Dutch mC4 🇳🇱

A GPT-Neo small model (125M paramters) trained from scratch on Dutch, with perplexity 19.9 on cleaned Dutch mC4.

How To Use

You can use this GPT-Neo model directly with a pipeline for text generation.

from transformers import pipeline, GPT2Tokenizer, GPTNeoForCausalLM
tokenizer = GPT2Tokenizer.from_pretrained(MODEL_DIR)
model = GPTNeoForCausalLM.from_pretrained(MODEL_DIR)
generator = pipeline('text-generation', model, tokenizer=tokenizer)

generated_text = generator('Wetenschappers verbonden aan de Katholieke Universiteit', max_length=100, do_sample=True, top_k=40, top_p=0.95, repetition_penalty=2.0))

"Wetenschappers verbonden aan de Katholieke Universiteit van Nijmegen" - "hebben er in het laatste nummer dat deze week verschijnt nog niets over gezegd. De wetenschappers verwachten pas volgend jaar meer duidelijkheid te kunnen geven, zo blijkt uit onderzoek door een vakblad en op Facebook onder studenten die denken mee te moeten werken om hun studie af te maken. In augustus 2017 kwam al naar buiten wat eraan schortten: hogescholen zouden moeite hebben met excel-software, ze hadden niet voldoende tijd om alle"


  • BPE tokenizer trained from scratch for Dutch on mC4 nl cleaned with scripts from the Huggingface Transformers Flax examples.


This model was trained on the wikipedia and newspapers (3.9B tokens) webpages in cleaned Dutch mC4, which is the original mC4, except

  • Documents that contained words from a selection of the Dutch and English List of Dirty Naught Obscene and Otherwise Bad Words are removed
  • Sentences with less than 3 words are removed
  • Sentences with a word of more than 1000 characters are removed
  • Documents with less than 5 sentences are removed
  • Documents with "javascript", "lorum ipsum", "terms of use", "privacy policy", "cookie policy", "uses cookies", "use of cookies", "use cookies", "elementen ontbreken", "deze printversie" are removed.


TL;DR: yhavinga/gpt2-medium-dutch is the best model.

  • yhavinga/gpt-neo-125M-dutch is trained on a fraction of C4 containing only wikipedia and news sites.
  • The models with a/b in the step-column have been trained to step a of a total of b steps.
model params train seq len ppl loss batch size epochs steps optim lr duration config
yhavinga/gpt-neo-125M-dutch gpt neo 125M 512 19.9 2.99 128 8 558608 adamw 2.4e-3 1d 12h news+wiki
yhavinga/gpt2-medium-dutch gpt2 345M 512 15.1 2.71 128 4 320000/520502 adafactor 8e-4 7d 2h full
yhavinga/gpt2-large-dutch gpt2 762M 512 15.1 2.72 32 1 1100000/2082009 adafactor 3.3e-5 8d 15h large
yhavinga/gpt-neo-1.3B-dutch gpt neo 1.3B 512 16.0 2.77 16 1 960000/3049896 adafactor 5e-4 7d 11h full


This project would not have been possible without compute generously provided by Google through the TPU Research Cloud. The HuggingFace 🤗 ecosystem was also instrumental in most, if not all, parts of the training. The following repositories where helpful in setting up the TPU-VM, and training the models:

Created by Yeb Havinga

Downloads last month

Dataset used to train yhavinga/gpt-neo-125M-dutch-nedd

Space using yhavinga/gpt-neo-125M-dutch-nedd 1