metadata

language:
  - nl
datasets:
  - yhavinga/mc4_nl_cleaned
tags:
  - seq2seq
  - lm-head
license: apache-2.0
inference: false

T5-base pre-trained on cleaned Dutch mC4 🇳🇱

A T5-base model trained from scratch on Dutch. This model is a re-training of the original t5-base-dutch model that was trained during the summer 2021 HuggingFace Flax/Jax community week. The original training suffered from errors during training, and was not trained on the full dataset. These errors have been fixed and training was now completed. Both this model and the flax-community t5-base-dutch model now have the same latest checkpoint with accuracy 0.70 and loss 1,38 on the validation split.

NB! Consider using yhavinga/t5-v1.1-base-dutch-cased that has an accuracy of 0,78 and loss 0,96 on the validation split.

These models need to be finetuned, therefore the inference widget on the right has been turned off. For a demo of the Dutch CNN summarization models, head over to the Hugging Face Spaces for the Netherformer 📰 example application!

Tokenizer

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

Dataset

All models listed below are trained on of the full configuration (39B tokens) of 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.

Models

The first model, t5-base-dutch is a re-training of the Dutch T5 base v1.0 model trained during the Flax/Jax community week. With training complete, accuracy was improved from 0,64 to 0,70.
The second two models are a uncased and cased version of t5-v1.1-base, again pre-trained from scratch on Dutch, with a tokenizer also trained from scratch. The t5 v1.1 models are slightly different from the t5 models, and the base models are trained with a dropout of 0.0. For fine-tuning it is intended to set this back to 0.1.
The large cased model is a pre-trained Dutch version of t5-v1.1-large. Training of t5-v1.1-large proved difficult. Without dropout regularization, the training would diverge at a certain point. With dropout training went better, be it much slower than training the t5-model. At some point convergance was too slow to warrant further training. The latest checkpoint, training scripts and metrics are available for reference. For actual fine-tuning the cased base model is probably the better choice.

	model	train seq len	acc	loss	batch size	epochs	steps	dropout	optim	lr	duration
t5-base-dutch	T5	512	0,70	1,38	128	1	528481	0.1	adafactor	5e-3	2d 9h
t5-v1.1-base-dutch-uncased	t5-v1.1	1024	0,73	1,20	64	2	1014525	0.0	adafactor	5e-3	5d 5h
t5-v1.1-base-dutch-cased	t5-v1.1	1024	0,78	0,96	64	2	1210000	0.0	adafactor	5e-3	6d 6h
t5-v1.1-large-dutch-cased	t5-v1.1	512	0,76	1,07	64	1	1120000	0.1	adafactor	5e-3	86 13h

The cased t5-v1.1 Dutch models were fine-tuned on summarizing the CNN Daily Mail dataset.

	model	input len	target len	Rouge1	Rouge2	RougeL	RougeLsum	Test Gen Len	epochs	batch size	steps	duration
t5-v1.1-base-dutch-cnn-test	t5-v1.1	1024	96	34,8	13,6	25,2	32,1	79	6	64	26916	2h 40m
t5-v1.1-large-dutch-cnn-test	t5-v1.1	1024	96	34,4	13,6	25,3	31,7	81	5	16	89720	11h

Acknowledgements

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 many, if not all parts of the training. The following repositories where helpful in setting up the TPU-VM, and getting an idea what sensible hyper-parameters are for training gpt2 from scratch.

Created by Yeb Havinga