06f1fd69a6
4.6 KiB
This model was published in HF papers on 2023-04-18 and contributed to Hugging Face Transformers on 2023-07-03.
UMT5
Overview
The UMT5 model was proposed in UniMax: Fairer and More Effective Language Sampling for Large-Scale Multilingual Pretraining by Hyung Won Chung, Xavier Garcia, Adam Roberts, Yi Tay, Orhan Firat, Sharan Narang, Noah Constant.
The abstract from the paper is the following:
Pretrained multilingual large language models have typically used heuristic temperature-based sampling to balance between different languages. However previous work has not systematically evaluated the efficacy of different pretraining language distributions across model scales. In this paper, we propose a new sampling method, UniMax, that delivers more uniform coverage of head languages while mitigating overfitting on tail languages by explicitly capping the number of repeats over each language's corpus. We perform an extensive series of ablations testing a range of sampling strategies on a suite of multilingual benchmarks, while varying model scale. We find that UniMax outperforms standard temperature-based sampling, and the benefits persist as scale increases. As part of our contribution, we release: (i) an improved and refreshed mC4 multilingual corpus consisting of 29 trillion characters across 107 languages, and (ii) a suite of pretrained umT5 model checkpoints trained with UniMax sampling.
Google has released the following variants:
This model was contributed by agemagician and stefan-it. The original code can be found here.
Usage tips
- UMT5 was only pre-trained on mC4 excluding any supervised training. Therefore, this model has to be fine-tuned before it is usable on a downstream task, unlike the original T5 model.
- Since umT5 was pre-trained in an unsupervised manner, there's no real advantage to using a task prefix during single-task fine-tuning. If you are doing multi-task fine-tuning, you should use a prefix.
Differences with mT5?
UmT5 is based on mT5, with a non-shared relative positional bias that is computed for each layer. This means that the model set has_relative_bias for each layer.
The conversion script is also different because the model was saved in t5x's latest checkpointing format.
Sample usage
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
model = AutoModelForSeq2SeqLM.from_pretrained("google/umt5-small", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("google/umt5-small")
inputs = tokenizer(
"A <extra_id_0> walks into a bar and orders a <extra_id_1> with <extra_id_2> pinch of <extra_id_3>.",
return_tensors="pt",
)
outputs = model.generate(**inputs)
print(tokenizer.batch_decode(outputs))
['<pad><extra_id_0>nyone who<extra_id_1> drink<extra_id_2> a<extra_id_3> alcohol<extra_id_4> A<extra_id_5> A. This<extra_id_6> I<extra_id_7><extra_id_52><extra_id_53></s>']
Refer to T5's documentation page for more tips, code examples and notebooks.
UMT5Config
autodoc UMT5Config
UMT5Model
autodoc UMT5Model - forward
UMT5ForConditionalGeneration
autodoc UMT5ForConditionalGeneration - forward
UMT5EncoderModel
autodoc UMT5EncoderModel - forward
UMT5ForSequenceClassification
autodoc UMT5ForSequenceClassification - forward
UMT5ForTokenClassification
autodoc UMT5ForTokenClassification - forward
UMT5ForQuestionAnswering
autodoc UMT5ForQuestionAnswering - forward