Models#
This page lists the available pre-trained T5 models. To use a pre-trained model, you need a Gin config file that defines the model params, and the model checkpoint to load from. For your convenience, TensorFlow checkpoints and Gin configs for common T5 pre-trained models have been made available for use in T5X. Following is a list of these pre-trained models and their Gin and checkpoint locations.
All checkpoints:
gs://t5-data/pretrained_models/t5x/All Gin files:
t5x/configs/models/
Selecting a model:#
Publicly Available Models:
Model |
Use Case |
|---|---|
Improved T5, recommended for most research. English only. |
|
The original T5 work for reproducibility. English only. |
|
Trained for 100k additional steps on the LM objective, per prompt tuning paper. |
|
Multilingual T5. Recommended for multilingual research. Note that at smaller scales (at least through XL), mT5 performance is lower than T5 on English tasks. |
|
Trained for 100k additional steps on the LM objective, per zero-shot cross-lingual generation (XGen) paper. |
|
umT5, an updated mT5 model trained using a more uniform language distribution, per the UniMax paper. |
|
ByT5. A “token-free” model that uses UTF-8 bytes for input and output. Recommended for tasks involving word-internal phenomena such as spelling, pronunciation, or morphology. |
|
Recommended checkpoints to fine-tune for long input sequence tasks |
|
Useful for MoE experimentation. |
|
General purpose T5 checkpoints for few-shot and finetuning. We recommend Flan-T5 over vanilla T5 and T5 LM-adapted |
|
Checkpoints for 20B pretrained and FLAN-based instruction-tuned models using the UL2 objective from UL2 paper |
|
Checkpoints from the BigScience paper |
|
Language-Image models trained with an alternative to CLIP, presented in the FLIP paper |
|
1.2B parameter encoder model for English to score model generations given a prefix for decoding from the RankGen paper |
|
11B parameter paraphrase generation model from the Dipper paper |
Public Research Models#
T5 Checkpoints#
These are the checkpoints used in the paper Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. They are encoder-decoder models pre-trained on C4 with a “span corruption” denoising objective, in addition to a mixture of downstream tasks including: GLUE, SuperGLUE, CNN/Daily Mail, SQuAD, and WMT.
Vocabulary: cc_all.32000.100extra
Model |
Gin File Location |
Checkpoint Location |
|---|---|---|
T5 Small |
gs://t5-data/pretrained_models/t5x/t5_small/checkpoint_1000000 |
|
T5 Base |
gs://t5-data/pretrained_models/t5x/t5_base/checkpoint_999900 |
|
T5 Large |
gs://t5-data/pretrained_models/t5x/t5_large/checkpoint_1000700 |
|
T5 3B |
||
T5 11B |
gs://t5-data/pretrained_models/t5x/t5_11B/checkpoint_1000000 |
T5 1.1 Checkpoints#
These are similar to the models from Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer, but with the following improvements:
GEGLU activation in feed-forward hidden layer, rather than ReLU - see https://arxiv.org/abs/2002.05202 .
Dropout was turned off in pre-training (quality win). Dropout should be re-enabled during fine-tuning.
Pre-trained on C4 only without mixing in the downstream tasks.
no parameter sharing between embedding and classifier layer
“xl” and “xxl” replace “3B” and “11B”. The model shapes are a bit different - larger d_model and smaller num_heads and d_ff.
For English-language, sequence-to-sequence-style tasks (ones where the goal is to map from an input text sequence to a target sequence) these are usually the best models to fine-tune.
Vocabulary: cc_all.32000.100extra
Model |
Gin File Location |
Checkpoint Location |
|---|---|---|
T5 1.1 Small |
gs://t5-data/pretrained_models/t5x/t5_1_1_small/checkpoint_1000000 |
|
T5 1.1 Base |
gs://t5-data/pretrained_models/t5x/t5_1_1_base/checkpoint_1000000 |
|
T5 1.1 Large |
gs://t5-data/pretrained_models/t5x/t5_1_1_large/checkpoint_1000000 |
|
T5 1.1 XL |
gs://t5-data/pretrained_models/t5x/t5_1_1_xl/checkpoint_1000000 |
|
T5 1.1 XXL |
gs://t5-data/pretrained_models/t5x/t5_1_1_xxl/checkpoint_1000000 |
T5 1.1 LM-Adapted Checkpoints#
These “LM-adapted” models are initialized from T5 1.1 (above) and trained for an additional 100K steps on the LM objective discussed in the T5 paper. This adaptation improves the ability of the model to be used for prompt tuning. These checkpoints were also used within the BigScience T0 project.
Vocabulary: cc_all.32000.100extra
Model |
Gin File Location |
Checkpoint Location |
|---|---|---|
T5 1.1 LM-100K Small |
||
T5 1.1 LM-100K Base |
||
T5 1.1 LM-100K Large |
||
T5 1.1 LM-100K XL |
||
T5 1.1 LM-100K XXL |
mT5 Checkpoints#
These are the checkpoints used in the paper mT5: A Massively Multilingual Pre-trained Text-to-Text Transformer. They are encoder-decoder models trained on multilingual C4 with a denoising objective. These are the best checkpoints to fine-tune for non-English sequence-to-sequence tasks.
Vocabulary: mc4.250000.100extra
Model |
Gin File Location |
Checkpoint Location |
|---|---|---|
mT5 Small |
gs://t5-data/pretrained_models/t5x/mt5_small/checkpoint_1000000 |
|
mT5 Base |
gs://t5-data/pretrained_models/t5x/mt5_base/checkpoint_1000000 |
|
mT5 Large |
gs://t5-data/pretrained_models/t5x/mt5_large/checkpoint_1000000 |
|
mT5 XL |
gs://t5-data/pretrained_models/t5x/mt5_xl/checkpoint_1000000 |
|
mT5 XXL |
gs://t5-data/pretrained_models/t5x/mt5_xxl/checkpoint_1000000 |
mT5 LM-Adapted Checkpoints#
These are the checkpoints released as part of the zero-shot cross-lingual generation (XGen) paper.
These “LM-adapted” models are initialized from mT5 (above) and trained for an additional 100K steps on the LM objective discussed in the T5 paper.
This adaptation improves the ability of the model to be used for prompt tuning.
Vocabulary: mc4.250000.100extra
Model |
Gin File Location |
Checkpoint Location |
|---|---|---|
mT5 LM-Adapted Small |
||
mT5 LM-Adapted Base |
||
mT5 LM-Adapted Large |
||
mT5 LM-Adapted XL |
||
mT5 LM-Adapted XXL |
umT5 Checkpoints#
These are the checkpoints described in the paper UniMax: Fairer and More Effective Language Sampling for Large-Scale Multilingual Pretraining. umT5 is similar to mT5 (see above); both are multilingual encoder-decoder models ranging from 300M to 13B parameters, trained on the mC4 corpus using a denoising objective. umT5 is trained on a fresher version of the mC4 corpus (3.1.0), and with a more uniform language balancing strategy.
Vocabulary: umt5.256000
Model |
Gin File Location |
Checkpoint Location |
|---|---|---|
umT5 Small |
||
umT5 Base |
||
umT5 XL |
||
umT5 XXL |
ByT5 Checkpoints#
These are the checkpoints used in the paper ByT5: Towards a Token-Free Future with Pre-trained Byte-to-Byte Models. They are similar to mT5 (above), but are “token-free”, processing text as raw UTF-8 bytes, as opposed to using a pretrained subword vocabulary. These models are more robust to character-level noise, and outperform parameter-matched mT5 models in many settings, particularly on word-level tasks sensitive to spelling, pronunciation, or morphology. However inference is significantly slower, up to 10x depending on the task.
Vocabulary: None
Model |
Gin File Location |
Checkpoint Location |
|---|---|---|
ByT5 Small |
gs://t5-data/pretrained_models/t5x/byt5_small/checkpoint_1000000 |
|
ByT5 Base |
gs://t5-data/pretrained_models/t5x/byt5_base/checkpoint_1000000 |
|
ByT5 Large |
gs://t5-data/pretrained_models/t5x/byt5_large/checkpoint_1000000 |
|
ByT5 XL |
gs://t5-data/pretrained_models/t5x/byt5_xl/checkpoint_1000000 |
|
ByT5 XXL |
gs://t5-data/pretrained_models/t5x/byt5_xxl/checkpoint_1000000 |
LongT5 Checkpoints#
These are the checkpoints used in the paper LongT5: Efficient Text-to-Text Transformer for Long Sequences. They are encoder-decoder models trained on C4 using the PEGASUS Principle Sentences Generation objective. These are the recommended checkpoints to fine-tune for long input sequence tasks.
LongT5 Local Attention Checkpoints#
The checkpoints below use local attention, which uses a sliding window to reduce training time from quadratic (with regards to input length) to linear. These are the recommended checkpoints to use for faster training/inference time.
Vocabulary: cc_all.32000.100extra
Model |
Gin File Location |
Checkpoint Location |
|---|---|---|
LongT5 Local Attention Base |
gs://t5-data/pretrained_models/t5x/longt5/local_base/checkpoint_1000000 |
|
LongT5 Local Attention Large |
gs://t5-data/pretrained_models/t5x/longt5/local_large/checkpoint_1000000 |
LongT5 Transient Global Attention Checkpoints#
The checkpoints below use transient global attention, which introduces global tokens at each encoder layer to allow tokens to interact with each other at longer distances. These are the recommended checkpoints to use for increased performance on long input sequence tasks.
Vocabulary: cc_all.32000.100extra
Model |
Gin File Location |
Checkpoint Location |
|---|---|---|
LongT5 Base |
gs://t5-data/pretrained_models/t5x/longt5/tglobal_base/checkpoint_1000000 |
|
LongT5 Large |
gs://t5-data/pretrained_models/t5x/longt5/tglobal_large/checkpoint_1000000 |
|
LongT5 XL |
gs://t5-data/pretrained_models/t5x/longt5/tglobal_xl/checkpoint_1000000 |
Mixture of Experts (MoE) Checkpoints#
These MoE checkpoints need to be used with T5X MoE overrides – specifically, the MoeTrainer and the MoePjitPartitioner. For example, for fine-tuning, use the MoE fine-tune run config.
Converted Mesh Tensorflow checkpoints#
These are the checkpoints from the Switch Transformer model.
Vocabulary: cc_all.32000.100extra
Flan-T5 Checkpoints#
These are the checkpoints released as part of the paper Scaling Instruction-Finetuned Language Models. They were initialized from the T5 1.1 LM-Adapted and instruction-finetuned.
They significantly outperform the LM-adapted checkpoints. For example, Flan-T5-XXL outperforms T5-LM-XXL by 26.6% absolute on the normalized average score. It even outperforms a much larger PaLM 62B model on BigBench Hard a set of challenging BigBench benchmark.
Unlike the vanilla T5 checkpoints, these can be directly used for few-shot prompting as well as standard finetuning. See Chung et al. 2022 for details.
Model |
Gin File Location |
Checkpoint Location |
|---|---|---|
Flan-T5 Small |
gs://t5-data/pretrained_models/t5x/flan_t5_small/checkpoint_1198000 |
|
Flan-T5 Base |
gs://t5-data/pretrained_models/t5x/flan_t5_base/checkpoint_1184000 |
|
Flan-T5 Large |
gs://t5-data/pretrained_models/t5x/flan_t5_large/checkpoint_1164000 |
|
Flan-T5 XL |
gs://t5-data/pretrained_models/t5x/flan_t5_xl/checkpoint_1138000 |
|
Flan-T5 XXL |
gs://t5-data/pretrained_models/t5x/flan_t5_xxl/checkpoint_1114000 |
UL2 Checkpoints#
Checkpoints for 20B pretrained and FLAN-based instruction-tuned models using the UL2 objective from UL2 paper. Checkpoints are released at https://github.com/google-research/google-research/tree/master/ul2#checkpoints.
BigScience Checkpoints#
Checkpoints from the BigScience paper, released at https://github.com/bigscience-workshop/architecture-objective/tree/main#checkpoints.
FLIP Checkpoints#
Language-Image models trained with an alternative to CLIP, presented in the FLIP paper. Checkpoints are released at https://github.com/facebookresearch/flip#results-and-pre-trained-flip-models.
RankGen Checkpoints#
1.2B parameter encoder model for English to score model generations given a prefix for decoding from the RankGen paper. Checkpoints are released at https://github.com/google-research/google-research/tree/master/rankgen.
Dipper Checkpoints#
11B parameter paraphrase generation model from the Dipper paper. Checkpoints are released at https://github.com/google-research/google-research/tree/master/dipper.