metadata

language:
  - nl
  - en
  - multilingual
license: apache-2.0
tags:
  - dutch
  - english
  - t5
  - t5x
  - ul2
  - seq2seq
  - translation
datasets:
  - yhavinga/mc4_nl_cleaned
  - yhavinga/nedd_wiki_news
pipeline_tag: translation
widget:
  - text: >-
      Redistricting and West Virginia’s shrinking population forced the state’s
      Republican Legislature to pit Mr. McKinley, a six-term Republican with a
      pragmatic bent, against Mr. Mooney, who has served four terms marked more
      by conservative rhetoric than legislative achievements.
  - text: >-
      It is a painful and tragic spectacle that rises before me: I have drawn
      back the curtain from the rottenness of man. This word, in my mouth, is at
      least free from one suspicion: that it involves a moral accusation against
      humanity.
  - text: >-
      Young Wehling was hunched in his chair, his head in his hand. He was so
      rumpled, so still and colorless as to be virtually invisible. His
      camouflage was perfect, since the waiting room had a disorderly and
      demoralized air, too. Chairs and ashtrays had been moved away from the
      walls. The floor was paved with spattered dropcloths.

ul2-base-nl36-en-nl for English to Dutch translation

Fine-tuned T5 model on English to Dutch translation that was pretrained on Dutch 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.

Model description

T5 is an encoder-decoder model and treats all NLP problems in a text-to-text format.

ul2-base-nl36-en-nl T5 is a transformers model fine-tuned on parallel sentence and paragraph pairs sampled from books.

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

The "efficient" T5 architecture findings presented in this paper were also applied, which suggests that a Deep-Narrow model architecture is favorable for downstream performance compared to other model architectures of similar parameter count. Specifically, the model depth is defined as the number of transformer blocks that are stacked sequentially. This model uses the t5-efficient-base-nl36 architecture's layer depth, which means both the encoder and the decoder have 36 transformer layers compared to the original T5 "base" model's architecture of 12 transformer layers.

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:

R-denoising (or regular span corruption), which emulates the standard T5 span corruption objective;
X-denoising (or extreme span corruption); and
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 fine-tuned on parallel sentence and paragraph pairs and can be used for machine translation.

How to use

Here is how to use this model in PyTorch:

model_name = "yhavinga/ul2-base-nl36-en-nl"
from transformers import AutoTokenizer
from transformers import AutoModelForSeq2SeqLM
from transformers import pipeline
import torch
device_num = 0 if torch.cuda.is_available() else -1
device = "cpu" if device_num < 0 else f"cuda:{device_num}"

tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name, use_auth_token=True).to(
    device
)
params = {"max_length": 370, "num_beams": 4, "early_stopping": True}
translator = pipeline("translation", tokenizer=tokenizer, model=model, device=device_num)
print(translator("Young Wehling was hunched in his chair, his head in his hand. He was so rumpled, so still and colorless as to be virtually invisible.",
               **params)[0]['translation_text'])

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-base-nl36-en-nl T5 model was pre-trained simultaneously on a combination of several datasets, including the full 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), 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.

After pre-training, the model was fine-tuned on a translation dataset containing 13 million sentence and paragraph pairs sampled from books.

Training procedure

Preprocessing

The ul2-base-nl36-en-nl 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.

Fine-tuning

This model was fine-tuned on a dataset containing 13M sentence and paragraph translation pairs sampled from books.

Pre-trained model used as starting point: yhavinga/ul2-base-nl36-dutch
Amount of fine-tune training steps: 43415
Batch size: 512 (gradient accumulation steps: 16)
Sequence length: 370 tokens
Model dtype: bfloat16
z_loss: 0.0001
Optimizer: adamw_hf beta1: 0.9 beta2: 0.9969 eps: 1e-08
Dropout rate: 0.01
Learning rate: 0.0009 with linear decay to 0 and warmup for 500 steps
Label smoothing factor: 0.11
Bleu score: 44.2

Model list

Models in this series:

	ul2-base-en-nl	ul2-base-nl36-en-nl	ul2-large-en-nl
model_type	t5	t5	t5
_pipeline_tag	translation	translation	translation
d_model	768	768	1024
d_ff	2048	3072	2816
num_heads	12	12	16
d_kv	64	64	64
num_layers	12	36	24
num_decoder_layers	12	36	24
feed_forward_proj	gated-silu	gated-silu	gated-silu
dense_act_fn	silu	silu	silu
vocab_size	32128	32128	32128
tie_word_embeddings	0	0	0
torch_dtype	float32	float32	float32
_gin_batch_size	128	64	64
_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