Team:

MERA

Name of the ML model:

Google's UMT5 XXL

Link to the ML model:

Additional links:

Architecture description:

Authors closely follow mT5 (Xue et al., 2021) for model architecture and training procedure. Specifically, thet use an encoder-decoder Transformer architecture and the span corruption pretraining objective from T5 (Raffel et al., 2020) on a multilingual corpus consisting of 101 languages plus 6 Latin-script variants (e.g. ru-Latn). They use batch size of 1024 sequences where each sequence is defined by selecting a chunk of 568 tokens from the training corpus. This is then split into 512 input and 114 target tokens. The number of training steps is 250,000.

Description of the training:

The model architectures used in this study are the same as mT5 models, except that relative position embeddings are not shared across layers. The vocabulary size is 256,000 subwords, and byte-level fallback is enabled, so unknown tokens are broken down into UTF-8 bytes. Authors use the T5X library (Roberts et al., 2022) to train the models using Google Cloud TPUs. For pretraining, they use Adafactor optimizer (Shazeer & Stern, 2018) with a constant learning rate of 0.01 in the first 10,000 steps and inverse square root decay afterwards. For finetuning, they use Adafactor with a constant learning rate of 5e−5.

Pretrain data:

UMT5 is pretrained on the an updated version of mC4 corpus, covering 107 languages.

Training Details:

Unlike mT5, authors do not use loss normalization factor. Instead they use the number of real target tokens as the effective loss normalization. Finally, they do not factorize the second moment of the Adafactor states and use momentum, neither of which are used in T5 and mT5 studies.

License:

Apache 2.0

Strategy, generation and parameters:

Code version v.1.1.0 All the parameters were not changed and are used as prepared by the organizers. Details: - 1 x NVIDIA A100 - dtype auto - Pytorch 2.1.2 + CUDA 12.1 - Transformers 4.36.2 - Context length 512