ByT5 is a tokenizer-free extension of the mT5 model. Instead of using a subword vocabulary like most other pretrained language models (BERT, XLM-R, T5, GPT-3), our ByT5 model operates directly on UTF-8 bytes, removing the need for any text preprocessing. Beyond the reduction in system complexity, we find that parameter-matched ByT5 models are competitive with mT5 across a range of tasks, and outperform mT5 on tasks that involve noisy text or are sensitive to spelling and pronunciation. This repo can be used to reproduce the experiments in the ByT5 paper.
To run this code, you need to install the t5
library. General instructions for training,
fine-tuning, evaluation, and exporting models for inference can be found in the
t5
repo. In
order to use the additional ByT5 tasks provided in this library with the
t5_mesh_transformer
command, run from this directory and add the flag
--module_import="byt5.tasks"
.
To train a ByT5-Large
model on the
mc4
task from scratch as described in the paper:
export PROJECT=yourproject
export ZONE=yourzone
export BUCKET=yourbucket
export TPU=yourtpu
ctpu up --name=$TPU --project=$PROJECT --zone=$ZONE --tpu-size=v3-256 --tpu-only --noconf
TASK=byt5_mc4
MODEL_DIR="${BUCKET}${TASK}"
python -m t5.models.mesh_transformer_main \
--tpu="${TPU}" \
--gcp_project="${PROJECT}" \
--tpu_zone="${ZONE}" \
--model_dir="${MODEL_DIR}" \
--gin_file="models/byt5.large.gin" \
--gin_param="MIXTURE_NAME = '${TASK}'" \
--gin_param="utils.run.sequence_length = {'inputs': 1024, 'targets': 189}" \
--gin_param="utils.run.batch_size = ('tokens_per_batch', 1048576)" \
--gin_param="utils.run.learning_rate_schedule=@learning_rate_schedules.rsqrt_no_ramp_down" \
--gin_param="run.train_steps = 1000000" \
--gin_param="utils.tpu_mesh_shape.model_parallelism = 1" \
--gin_param="utils.tpu_mesh_shape.tpu_topology = 'v3-256'" \
--eval_mode="perplexity_eval" \
--eval_gin_param="mesh_eval_dataset_fn.num_eval_examples = 10000" \
--t5_tfds_data_dir="${BUCKET}/t5-tfds" \
--module_import="byt5.tasks"
The example below shows how to finetune the ByT5-Large
model on the XNLI
zeroshot task.
export PROJECT=yourproject
export ZONE=yourzone
export BUCKET=yourbucket
export TPU=yourtpu
ctpu up --name=$TPU --project=$PROJECT --zone=$ZONE --tpu-size=v3-256 --tpu-only --noconf
TASK=byt5_xnli_zeroshot
PRETRAINED_DIR=gs:https://t5-data/pretrained_models/byt5/large
PRETRAINED_STEPS=1000000
FINETUNE_STEPS=262144
MODEL_DIR="${BUCKET}${TASK}"
# Run fine-tuning
python -m t5.models.mesh_transformer_main \
--tpu="${TPU}" \
--gcp_project="${PROJECT}" \
--tpu_zone="${ZONE}" \
--model_dir="${MODEL_DIR}" \
--gin_file="${PRETRAINED_DIR}/operative_config.gin" \
--gin_param="utils.tpu_mesh_shape.tpu_topology = 'v3-256'" \
--gin_param="MIXTURE_NAME = '${TASK}'" \
--gin_param="utils.run.train_steps=$((PRETRAINED_STEPS+FINETUNE_STEPS))" \
--gin_param="utils.run.init_checkpoint='${PRETRAINED_DIR}/model.ckpt-${PRETRAINED_STEPS}'" \
--t5_tfds_data_dir="${BUCKET}/t5-tfds" \
--module_import="byt5.tasks"
--gin_param="utils.run.batch_size = ('tokens_per_batch', 1048576)" \
--gin_param="utils.run.sequence_length = {'inputs': 2048, 'targets': 56}"
--eval_gin_param="Bitransformer.decode.max_decode_length = 56" \
The remaining experiments are shown in the tasks.py file.
We have released the following checkpoints for pre-trained models described in our paper:
- ByT5-Small (300 million parameters): gs:https://t5-data/pretrained_models/byt5/small
- ByT5-Base (580 million parameters): gs:https://t5-data/pretrained_models/byt5/base
- ByT5-Large (1.2 billion parameters): gs:https://t5-data/pretrained_models/byt5/large
- ByT5-XL (3.7 billion parameters): gs:https://t5-data/pretrained_models/byt5/xl
- ByT5-XXL (13 billion parameters): gs:https://t5-data/pretrained_models/byt5/xxl
If you extend or use this work, please cite the paper where it was introduced:
@misc{xue2021byt5,
title={ByT5: Towards a token-free future with pre-trained byte-to-byte models},
author={Linting Xue and Aditya Barua and Noah Constant and Rami Al-Rfou and Sharan Narang and Mihir Kale and Adam Roberts and Colin Raffel},
year={2021},
eprint={2105.13626},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
This is not an officially supported Google product.