Multilingual T5 (mT5) is a massively multilingual pretrained text-to-text transformer model, trained following a similar recipe as T5. This repo can be used to reproduce the experiments in the mT5 paper.

• Languages covered
• Results
• Usage
  • Training
  • Fine-Tuning
• Released Model Checkpoints
• How to Cite

mT5 is pretrained on the mC4 corpus, covering 101 languages:

Afrikaans, Albanian, Amharic, Arabic, Armenian, Azerbaijani, Basque, Belarusian, Bengali, Bulgarian, Burmese, Catalan, Cebuano, Chichewa, Chinese, Corsican, Czech, Danish, Dutch, English, Esperanto, Estonian, Filipino, Finnish, French, Galician, Georgian, German, Greek, Gujarati, Haitian Creole, Hausa, Hawaiian, Hebrew, Hindi, Hmong, Hungarian, Icelandic, Igbo, Indonesian, Irish, Italian, Japanese, Javanese, Kannada, Kazakh, Khmer, Korean, Kurdish, Kyrgyz, Lao, Latin, Latvian, Lithuanian, Luxembourgish, Macedonian, Malagasy, Malay, Malayalam, Maltese, Maori, Marathi, Mongolian, Nepali, Norwegian, Pashto, Persian, Polish, Portuguese, Punjabi, Romanian, Russian, Samoan, Scottish Gaelic, Serbian, Shona, Sindhi, Sinhala, Slovak, Slovenian, Somali, Sotho, Spanish, Sundanese, Swahili, Swedish, Tajik, Tamil, Telugu, Thai, Turkish, Ukrainian, Urdu, Uzbek, Vietnamese, Welsh, West Frisian, Xhosa, Yiddish, Yoruba, Zulu.

mT5 achieves state-of-the-art performance on many cross-lingual NLP tasks, as of November 2020. For example, on XTREME zero-shot classification, structured prediction and QA tasks (showing F1 scores):

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 mT5 tasks provided in this library with the t5_mesh_transformer command, run from this directory and add the flag --module_import="multilingual_t5.tasks". There is also support for mT5 in HuggingFace; see instructions in the T5 repo here.

To train an mT5-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=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/t5.1.1.large.gin" \
  --gin_param="MIXTURE_NAME = '${TASK}'" \
  --gin_param="utils.run.sequence_length = {'inputs': 1024, 'targets': 256}" \
  --gin_param="utils.run.batch_size = ('tokens_per_batch', 1048576)" \
  --gin_param="utils.run.learning_rate_schedule=@learning_rate_schedules.rsqrt_no_ramp_down" \
  --gin_param="run.train_steps = 1000000" \
  --gin_param="utils.tpu_mesh_shape.model_parallelism = 1" \
  --gin_param="utils.tpu_mesh_shape.tpu_topology = 'v3-256'" \
  --eval_mode="perplexity_eval" \
  --eval_gin_param="mesh_eval_dataset_fn.num_eval_examples = 10000" \
  --t5_tfds_data_dir="${BUCKET}/t5-tfds" \
  --module_import="multilingual_t5.tasks"

The example below shows how to finetune the mT5-Large model on the XNLI zeroshot task. See finetune_mt5_tasks.sh for hyperparameter settings for other tasks.

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=mt5_xnli_zeroshot
SEQUENCE_LENGTH_GIN=xnli
PRETRAINED_DIR=gs://t5-data/pretrained_models/mt5/large
PRETRAINED_STEPS=1000000
FINETUNE_STEPS=20000
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="${PRETRAINED_DIR}/operative_config.gin" \
  --gin_file="sequence_lengths/${SEQUENCE_LENGTH_GIN}.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="multilingual_t5.tasks" \
  --gin_param="utils.run.batch_size = ('tokens_per_batch', 1048576)" \
  --gin_location_prefix="multilingual_t5/gin/"

The remaining experiments are shown in the tasks.py file.

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

• mT5-Small (300 million parameters): gs://t5-data/pretrained_models/mt5/small
• mT5-Base (580 million parameters): gs://t5-data/pretrained_models/mt5/base
• mT5-Large (1.2 billion parameters): gs://t5-data/pretrained_models/mt5/large
• mT5-XL (3.7 billion parameters): gs://t5-data/pretrained_models/mt5/xl
• mT5-XXL (13 billion parameters): gs://t5-data/pretrained_models/mt5/xxl

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

@inproceedings{xue-etal-2021-mt5,
    title = "m{T}5: A Massively Multilingual Pre-trained Text-to-Text Transformer",
    author = "Xue, Linting  and
      Constant, Noah  and
      Roberts, Adam  and
      Kale, Mihir  and
      Al-Rfou, Rami  and
      Siddhant, Aditya  and
      Barua, Aditya  and
      Raffel, Colin",
    booktitle = "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jun,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.naacl-main.41",
    doi = "10.18653/v1/2021.naacl-main.41",
    pages = "483--498"
}