# ul2-large-dutch-english for Dutch and English

Pretrained T5 model on Dutch and English using a UL2 (Mixture-of-Denoisers) objective. The T5 model was introduced in this paper and first released at this page. The UL2 objective was introduced in this paper and first released at this page.

Note: The Hugging Face inference widget is deactivated because this model needs a text-to-text fine-tuning on a specific downstream task to be useful in practice.

## Model description

T5 is an encoder-decoder model and treats all NLP problems in a text-to-text format. ul2-large-dutch-english T5 is a transformers model pretrained on a very large corpus of Dutch and English data in a self-supervised fashion. This means it was pretrained on the raw texts only, with no humans labelling them in any way (which is why it can use lots of publicly available data) with an automatic process to generate inputs and outputs from those texts.

This model used the T5 v1.1 improvements compared to the original T5 model during the pretraining:

• GEGLU activation in the feed-forward hidden layer, rather than ReLU - see here
• Dropout was turned off during pre-training. Dropout should be re-enabled during fine-tuning
• Pre-trained on self-supervised objective only without mixing in the downstream tasks
• No parameter sharing between embedding and classifier layer

### UL2 pretraining objective

This model was pretrained with the UL2's Mixture-of-Denoisers (MoD) objective, that combines diverse pre-training paradigms together. UL2 frames different objective functions for training language models as denoising tasks, where the model has to recover missing sub-sequences of a given input. During pre-training it uses a novel mixture-of-denoisers that samples from a varied set of such objectives, each with different configurations. UL2 is trained using a mixture of three denoising tasks:

1. R-denoising (or regular span corruption), which emulates the standard T5 span corruption objective;
2. X-denoising (or extreme span corruption); and
3. S-denoising (or sequential PrefixLM).

During pre-training, we sample from the available denoising tasks based on user-specified ratios. UL2 introduces a notion of mode switching, wherein downstream fine-tuning is associated with specific pre-training denoising task. During the pre-training, a paradigm token is inserted to the input ([NLU] for R-denoising, [NLG] for X-denoising, or [S2S] for S-denoising) indicating the denoising task at hand. Then, during fine-tuning the same input token should be inserted to get the best performance for different downstream fine-tuning tasks.

## Intended uses & limitations

This model was only pretrained in a self-supervised way excluding any supervised training. Therefore, this model has to be fine-tuned before it is usable on a downstream task, like text classification, unlike the Google's original T5 model.

Note: You most likely need to fine-tune these T5/UL2 models without mixed precision so fine-tune them with full fp32 precision. Fine-tuning with Flax in bf16 - model.to_bf16() - is possible if you set the mask correctly to exclude layernorm and embedding layers. Also note that the T5x pre-training and fine-tuning configs set z_loss to 1e-4, which is used to keep the loss scale from underflowing. You can also find more fine-tuning tips from here, for example.

Note: For fine-tuning, most likely you can get better results if you insert a prefix token of [NLU], [NLG], or [S2S] to your input texts. For general language understanding fine-tuning tasks, you could use the [NLU] token. For GPT-style causal language generation, you could use the [S2S] token. The token [NLG] of the X-denoising pretrain task is somewhat mix between the language understanding and causal language generation so the token [NLG] could maybe be used for language generation fine-tuning too.

### How to use

Here is how to use this model in PyTorch:

from transformers import T5Tokenizer, T5ForConditionalGeneration

tokenizer = T5Tokenizer.from_pretrained("yhavinga/ul2-large-dutch-english", use_fast=False)
model = T5ForConditionalGeneration.from_pretrained("yhavinga/ul2-large-dutch-english")


and in Flax:

from transformers import T5Tokenizer, FlaxT5ForConditionalGeneration

tokenizer = T5Tokenizer.from_pretrained("yhavinga/ul2-large-dutch-english", use_fast=False)
model = FlaxT5ForConditionalGeneration.from_pretrained("yhavinga/ul2-large-dutch-english")


### Limitations and bias

The training data used for this model contains a lot of unfiltered content from the internet, which is far from neutral. Therefore, the model can have biased predictions. This bias will also affect all fine-tuned versions of this model.

## Training data

The ul2-large-dutch-english T5 model was pre-trained simultaneously on a combination of several datasets, including the full_en_nl config of the "mc4_nl_cleaned" dataset, which is a cleaned version of Common Crawl's web crawl corpus, Dutch books, the Dutch subset of Wikipedia (2022-03-20), the English subset of Wikipedia (2022-03-01), and a subset of "mc4_nl_cleaned" containing only texts from Dutch and Belgian newspapers. This last dataset is oversampled to bias the model towards descriptions of events in the Netherlands and Belgium.

## Training procedure

### Preprocessing

The ul2-large-dutch-english T5 model uses a SentencePiece unigram tokenizer with a vocabulary of 32,000 tokens. The tokenizer includes the special tokens <pad>, </s>, <unk>, known from the original T5 paper, [NLU], [NLG] and [S2S] for the MoD pre-training, and <n> for newline. During pre-training with the UL2 objective, input and output sequences consist of 512 consecutive tokens. The tokenizer does not lowercase texts and is therefore case-sensitive; it distinguises between dutch and Dutch. Additionally, 100+28 extra tokens were added for pre-training tasks, resulting in a total of 32,128 tokens.

### Pretraining

The model was trained on TPUv3-8 VM, sponsored by the Google TPU Research Cloud, for 1000000 steps with a batch size of 64 (in total 32 B tokens). The optimizer used was AdaFactor with learning rate warmup for 10K steps with a constant learning rate of 1e-2, and then an inverse square root decay (exponential decay) of the learning rate after. The model was trained with Google's Jax/Flax based t5x framework with help from Stephenn Fernandes to get started writing task definitions that wrap HF datasets.

The UL2 training objective code used with the t5x framework was copied and slightly modified from the UL2 paper appendix chapter 9.2 by the authors of the Finnish ul2 models. Used UL2 objective code is available in the repository Finnish-NLP/ul2-base-nl36-finnish in the files ul2_objective.py and tasks.py. UL2's mixture-of-denoisers configuration was otherwise equal to the UL2 paper but for the rate of mixing denoisers, 20% for S-denoising was used (suggested at the paper chapter 4.5) and the rest was divided equally between the R-denoising and X-denoising (i.e. 40% for both).

### Model list

Models in this series:

ul2-base-dutch-english ul2-large-dutch-english ul2-small-dutch-english
model_type t5 t5 t5
_pipeline_tag text2text-generation text2text-generation text2text-generation
d_model 768 1024 512
d_ff 2048 2816 1024
d_kv 64 64 64
num_layers 12 24 8
num_decoder_layers 12 24 8
feed_forward_proj gated-gelu gated-gelu gated-gelu
dense_act_fn gelu_new gelu_new gelu_new
vocab_size 32128 32128 32128
tie_word_embeddings 0 0 0
torch_dtype float32 float32 float32
_gin_batch_size 128 64 128
_gin_z_loss 0.0001 0.0001 0.0001
_gin_t5_config_dtype 'bfloat16' 'bfloat16' 'bfloat16'

## Evaluation results

See the evaluation section in the interactive Pre-training Dutch T5 Models blog.

## Acknowledgements

This project would not have been possible without compute generously provided by Google through the TPU Research Cloud. Thanks to the Finnish-NLP authors for releasing their code for the UL2 objective and associated task definitions. Thanks to Stephenn Fernandes for helping me get started with the t5x framework.

Created by Yeb Havinga