February 6, 2020

2275 words 11 mins read



Code for the paper “Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer”

repo name google-research/text-to-text-transfer-transformer
repo link https://github.com/google-research/text-to-text-transfer-transformer
homepage https://arxiv.org/abs/1910.10683
language Python
size (curr.) 404 kB
stars (curr.) 2016
created 2019-10-17
license Apache License 2.0

T5: Text-To-Text Transfer Transformer

Build Status

T5 serves primarily as code for reproducing the experiments in Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. In the paper, we demonstrate how to achieve state-of-the-art results on multiple NLP tasks using a text-to-text transformer pre-trained on a large text corpus.

The bulk of the code in this repository is used for loading, preprocessing, mixing, and evaluating datasets. It also provides a way to fine-tune the pre-trained models released alongside the publication.

T5 can be used as a library for future model development by providing useful modules for training and fine-tuning (potentially huge) models on mixtures of text-to-text tasks.

Table of Contents



t5.data is a package for defining Task objects that provide tf.data.Datasets.

Each Task is made up of:

  • a data source
  • text preprocessor function(s)
  • a SentencePiece model
  • metric function(s)

Additionally, you may optionally provide:

  • token preprocessor function(s)
  • a postprocess function

The data source can be an arbitrary function that provides a tf.data.Dataset, but we also provide simpler wrappers for datasets available in TensorFlow Datasets (TFDS) (a TfdsTask) or stored as text files with one example per line (a TextLineTask).

The text preprocessor converts the examples in the source dataset into the appropriate format for a text-to-text model with fields for inputs and targets. For example, the predefined t5.data.preprocessors.translate preprocessor converts inputs in the form

{'de': 'Das ist gut.', 'en': 'That is good.'}

to the form

{'inputs': 'translate German to English: Das ist gut.', 'targets': 'That is good.'}

In additon to text preprocessing, you can also use one or more token preprocessors to modify the inputs post-tokenization. We implemented our unsupervised pre-training objectives using these token preprocessors.

We provide many predefined preprocessors in t5.data.preprocessors, but you may also define your own.

The SentencePiece model is used to tokenize the input strings and decode the output tokens. You can create your own model with the google/sentencepiece library, or use our default one at t5.data.DEFAULT_SPM_PATH.

The metric function returns a score given the target and prediction from the model. You may also define a postprocess function to convert the target and prediction text to another format before calling the metric. We provide some predefined metrics in t5.evaluation.metrics.

Finally, t5.data contains a Mixture class that can be instantiated to combine multiple Task datasets for multi-task training using various functions for specifying the mixture rates.


t5.evaluation contains two core components:

  1. metrics to be used during evaluation
  2. utilities for applying these metrics at evaluation time


t5.models contains shims for connecting T5 Tasks and Mixtures to a model implementation for training, evaluation, and inference.

Currently the only available shim is to Mesh TensorFlow Transformer (MeshTF), which enables both data and model parallelism for training massive Transformer models. This also includes a binary for launching the model in MeshTF along with gin configs for setting various hyperparameters.


The easiest way to try out T5 is with a free TPU in our Colab Tutorial.

Below we provide examples for how to pre-train, fine-tune, evaluate, and decode from a model from the command-line with our codebase. You can use these instructions to reproduce our results, fine-tune one of our released checkpoints with your own data and/or hyperparameters, or pre-train a model from scratch.

Dataset Preparation

You may either use a new or pre-existing Task, or you may load examples from a preprocessed TSV file.

Using a Task

Depending on your data source (see above), you will need to prepare your data appropriately.


If using a vanilla task, just make sure any file(s) loaded by your dataset_fn are accessible to the TPU (i.e., are in a GCS bucket), and you should be good to go!


Most of our predefined Tasks use TensorFlow Datasets (TFDS) as their data source. When you run our training binary (see instructions below) with a TfdsTask, the dataset will automatically be downloaded and prepared on its first use. After preparation is complete, the dataset is cached to your local storage to avoid this overhead in future runs. If working in the cloud, we recommend you set the --t5_tfds_data_dir flag to point to a persistent storage location, such as a GCS bucket. This is a requirement when training on TPU.

Note:The C4 dataset we created for unsupervised pre-training is available in TensorFlow Datasets, but it requires a significant amount of bandwith for downloading the raw Common Crawl scrapes and compute for its preparation. We suggest you take advantage of the Apache Beam support in TFDS, which enables distributed preprocessing of the dataset and can be run on Google Cloud Dataflow. Otherwise, it is unlikely that you will be able to complete preprocessing in a human lifetime. Read more in the TFDS Beam instructions.


A TextLineTask is useful when your data source is a text file (or files) with one example per line. You can then use a text preprocessor to convert each line into a dictionary of inputs and targets.

Make sure your files are accessible to the TPU (i.e., are in a GCS bucket), and you should be good to go!

Using a TSV File Directly

Instead of defining a new Task, you may use a TSV file (or files) directly as your dataset where each line is formatted as <input>\t<target>.

However, there are a couple of caveats:

  • There is no way to define a text processor, so the TSV will need to contain your data in a preprocessed format.
  • There is also currently no way to set a token preprocessor, postprocess function, or metric function for evaluation when using a TSV file directly.

If you need any of these features, you must define a new Task, TfdsTask, or TextLineTask.

Similar to the above cases, your TSV file(s) must be accessible to the TPU (i.e., are in a GCS bucket).


To install the T5 package, simply run:

pip install t5[gcp]

Setting up TPUs on GCP

You will first need to launch a Virtual Machine (VM) on Google Cloud. Details about launching the VM can be found at the Google Cloud Documentation.

In order to run training or eval on Cloud TPUs, you must set up the following variables based on your project, zone and GCS bucket appropriately. Please refer to the Cloud TPU Quickstart guide for more details.

export PROJECT=your_project_name
export ZONE=your_project_zone
export BUCKET=gs://yourbucket/
export TPU_NAME=t5-tpu
export DATA_DIR="${BUCKET}/your_data_dir"
export MODEL_DIR="${BUCKET}/your_model_dir"

Please use the following command to create a TPU device in the Cloud VM.

ctpu up --name=$TPU_NAME --project=$PROJECT --zone=$ZONE --tpu-size=v3-8  \
        --tpu-only   --tf-version=1.15 --noconf


In the command below, we train a model on the GLUE Benchmark MRPC task from scratch. You can change the MIXTURE_NAME gin parameter to use any of the tasks or mixtures provided in our package.

t5_mesh_transformer  \
  --tpu="${TPU_NAME}" \
  --gcp_project="${PROJECT}" \
  --tpu_zone="${ZONE}" \
  --model_dir="${MODEL_DIR}" \
  --t5_tfds_data_dir="${DATA_DIR}" \
  --gin_file="dataset.gin" \
  --gin_file="models/bi_v1.gin" \
  --gin_param="utils.tpu_mesh_shape.model_parallelism = 1" \
  --gin_param="utils.tpu_mesh_shape.tpu_topology = '2x2'" \
  --gin_param="MIXTURE_NAME = 'glue_mrpc_v002'"

The full list of tasks and mixtures can be obtained by running:

python -c "import t5; print(t5.data.MixtureRegistry.names())"

You may also define additional tasks and mixtures in a new file and import it using the --module_import flag.

Alternatively, you could train with a TSV file where each line is formatted as <input>\t<target> (see above).


In order to fine-tune one of our pre-trained models, you need to pass the operative config of the pre-trained model to the training script. The operative config should be passed in as a gin_file flag. It specifies the model architecture and other hyperparameters. In addition, you need to specify the mixture to fine-tune on. For example, to fine-tune the T5-small model on the glue_mrpc_v002 mixture, please run:

t5_mesh_transformer  \
  --tpu="${TPU_NAME}" \
  --gcp_project="${PROJECT}" \
  --tpu_zone="${ZONE}" \
  --model_dir="${MODEL_DIR}" \
  --t5_tfds_data_dir="${DATA_DIR}" \
  --gin_file="dataset.gin" \
  --gin_param="utils.tpu_mesh_shape.model_parallelism = 1" \
  --gin_param="utils.tpu_mesh_shape.tpu_topology = '2x2'" \
  --gin_param="MIXTURE_NAME = 'glue_mrpc_v002'" \

The correct pre-trained checkpoint path is included in the operative config.

You may also define additional tasks and mixtures in a new file and import it using the --module_import flag.

Alternatively, you could fine-tune with a TSV file where each line is formatted as <input>\t<target> (see above). For example, you could try one of the paired translation datasets from WMT ‘19 News Commentary 14 training set (e.g., English-French). When using a TSV file, you would replace the MIXTURE_NAME flag with:

--gin_param="utils.run.train_dataset_fn = @t5.models.mesh_transformer.tsv_dataset_fn"
--gin_param="tsv_dataset_fn.filename = 'gs:/path/to/tsv'"

To fine-tune with the same hyperparameters we used in the paper (using a constant learning rate of 0.001), you can pass in this gin file which is included in the T5 package:


The operative config for the pre-trained models are set so that there is effectively no limit on the number of train steps. If you’d like to train for a specific number of steps, you’ll need to pass that in. Since the pre-trained model has already been trained for 1,000,000 steps, you should specify the total number of steps after pre-training and fine-tuning. For example, if you want to fine-tune for an additional 10,000 steps, you should pass

--gin_param="run.train_steps = 1010000"

You can also use a different batch size for fine-tuning. We set the batch size according to the total number of tokens in a batch. By default, a batch uses a sequence length of 512. To set the number of tokens in a batch, you should set

--gin_param = "tokens_per_batch=1048576"


In order to evaluate a model in the T5 framework, you need to use the eval.gin file, specify the model directory, decoding method, and which checkpoint step(s) to evaluate. So, to evaluate on the GLUE MRPC task using beam search on all checkpoints, use the following command:

t5_mesh_transformer \
  --tpu="${TPU_NAME}" \
  --gcp_project="${PROJECT}" \
  --tpu_zone="${ZONE}" \
  --model_dir="${MODEL_DIR}" \
  --gin_file="${MODEL_DIR}/operative_config.gin" \
  --t5_tfds_data_dir=${DATA_DIR} \
  --gin_file="eval.gin" \
  --gin_file="beam_search.gin" \
  --gin_param="utils.tpu_mesh_shape.tpu_topology = '2x2'" \
  --gin_param="MIXTURE_NAME = 'glue_mrpc_v002'" \
  --gin_param="eval_checkpoint_step = 'all'"

To evaluate a specific checkpoint, simply set the eval_checkpoint_step parameter to appropriate checkpoint.

--gin_param="eval_checkpoint_step = 100000"

You can also use greedy_decode.gin or sample_decode.gin instead of beam_search.gin in the command above.


In order to produce predictions from a model in the T5 framework, you need to specify the model directory, decoding method, and which checkpoint step(s) to use for decoding. Assuming you have a text file of input sequences stored at /path/to/intputs.txt, an example command would be:

t5_mesh_transformer \
  --tpu="${TPU_NAME}" \
  --gcp_project="${PROJECT}" \
  --tpu_zone="${ZONE}" \
  --model_dir="${MODEL_DIR}" \
  --gin_file="${MODEL_DIR}/operative_config.gin" \
  --gin_file="infer.gin" \
  --gin_file="sample_decode.gin" \
  --gin_param="input_filename = '/path/to/inputs.txt'"\
  --gin_param="output_filename = '/tmp/outputs.txt'"\
  --gin_param="utils.tpu_mesh_shape.tpu_topology = '2x2'"\
  --gin_param="infer_checkpoint_step = 'all'"

To predict with a specific checkpoint, simply set the infer_checkpoint_step parameter to appropriate checkpoint.

--gin_param="infer_checkpoint_step = 100000"

You can also use beam_search.gin or greedy_decode.gin instead of sample_decode.gin in the command above.


You may also want to export a SavedModel, which is useful for serving your trained model, (e.g., when deploying with ML Engine).

t5_mesh_transformer \
  --gcp_project="${PROJECT}" \
  --tpu_zone="${ZONE}" \
  --model_dir="${MODEL_DIR}" \
  --use_model_api \
  --mode="export" \

GPU Usage

If you would like to use GPU instead of TPUs, you can modify the above commands by removing TPU-specific flags (--tpu, --tpu_zone, --gcp_project) and setting the gin params for mesh_shape and mesh_devices based on your desired setup.

For example, if your machine has access to 6 GPUs and you’d like to do 3-way model parallelism and 2-way data parallelism, the fine-tuning command above would become:

t5_mesh_transformer  \
  --model_dir="${MODEL_DIR}" \
  --t5_tfds_data_dir="${DATA_DIR}" \
  --gin_file="dataset.gin" \
  --gin_param="utils.run.mesh_shape = 'model:3,batch:2'" \
  --gin_param="utils.run.mesh_devices = ['gpu:0','gpu:1','gpu:2','gpu:3','gpu:4','gpu:5']" \
  --gin_param="MIXTURE_NAME = 'glue_mrpc_v002'" \

With a single GPU, the command is:

t5_mesh_transformer  \
  --model_dir="${MODEL_DIR}" \
  --t5_tfds_data_dir="${DATA_DIR}" \
  --gin_file="dataset.gin" \
  --gin_param="utils.run.mesh_shape = 'model:1,batch:1'" \
  --gin_param="utils.run.mesh_devices = ['gpu:0']" \
  --gin_param="MIXTURE_NAME = 'glue_mrpc_v002'" \

Reproducing our experiments

We provide operative configs for all of the experiments in the paper in gs://t5-data/experiments. The experiments folder has different subdirectories corresponding to the different sections in our paper. For example, gs://t5-data/experiments/objectives contains the experiments from Section 3.3 (“Unsupervised objectives”). Each subdirectory of the objectives folder contains operative configs for some particular experiment (where loosely speaking an “experiment” is one of the rows in one of the tables in our paper).

Let’s say you want to reproduce the results for the “Prefix language modeling” objective (the first row in Table 4). The operative configs for that experiment live in gs://t5-data/experiments/objectives/obj-prefix_lm. In the base directory, there is an operative config for pre-training the model (gs://t5-data/experiments/objectives/obj-prefix_lm/operative_config.gin). Then, there are subdirectories for each of the downstream fine-tuning mixtures we consider, each of which has its own operative config (for example, gs://t5-data/experiments/objectives/obj-prefix_lm/cnn_dailymail_v002/operative_config.gin). To run this experiment, first pre-train a model with the pre-training operative config:

export PRETRAIN_MODEL_DIR="${BUCKET}/obj-prefix_lm"
t5_mesh_transformer  \
  --tpu="${TPU_NAME}" \
  --gcp_project="${PROJECT}" \
  --tpu_zone="${ZONE}" \
  --model_dir="${PRETRAIN_MODEL_DIR}" \
  --gin_file="gs://t5-data/experiments/objectives/obj-prefix_lm/operative_config.gin" \
  --gin_param="utils.tpu_mesh_shape.model_parallelism = 1" \
  --gin_param="utils.tpu_mesh_shape.tpu_topology = '2x2'"

Then, you can fine-tune the pre-trained model on CNN/Daily Mail like so:

export FINETUNE_MODEL_DIR="${BUCKET}/obj-prefix_lm/cnn_dailymail_v002"
t5_mesh_transformer  \
  --tpu="${TPU_NAME}" \
  --gcp_project="${PROJECT}" \
  --tpu_zone="${ZONE}" \
  --model_dir="${FINETUNE_MODEL_DIR}" \
  --gin_file="gs://t5-data/experiments/objectives/obj-prefix_lm/cnn_dailymail_v002/operative_config.gin" \
  --gin_param="init_checkpoint = '${PRETRAIN_MODEL_DIR}/model.ckpt-524288'" \
  --gin_param="utils.tpu_mesh_shape.model_parallelism = 1" \
  --gin_param="utils.tpu_mesh_shape.tpu_topology = '2x2'"

Released Model Checkpoints

We have released the following checkpoints for pre-trained models described in our paper:

How to Cite

If you extend or use this work, please cite the paper where it was introduced:

  author = {Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu},
  title = {Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer},
  journal = {arXiv e-prints},
  year = {2019},
  archivePrefix = {arXiv},
  eprint = {1910.10683},
comments powered by Disqus