View source for T5 (language model)

{{Short description|Series of large language models developed by Google AI}}
{{Infobox software
| name = Text-to-Text Transfer Transformer (T5)
| logo = 
| screenshot = 
| screenshot size = 
| caption = 
| author = [[Google AI]]
| developer = 
| released = {{Start date and age|23 October 2019}}
| latest release version = T5X {{url|https://github.com/google-research/t5x}}
| latest release date = 
| repo = https://github.com/google-research/text-to-text-transfer-transformer
| programming language = 
| operating system = 
| replaces = 
| replaced_by = 
| genre = {{ indented plainlist |
*[[Large language model]]
*[[Transformer (deep learning architecture)]]
}}
| license = [[Apache-2.0]]
| website = {{url|https://blog.research.google/2020/02/exploring-transfer-learning-with-t5.html}}
}}

'''T5 (Text-to-Text Transfer Transformer)''' is a series of [[Large language model|large language models]] developed by [[Google AI]] introduced in 2019.<ref name=":0">{{Cite journal |last1=Raffel |first1=Colin |last2=Shazeer |first2=Noam |last3=Roberts |first3=Adam |last4=Lee |first4=Katherine |last5=Narang |first5=Sharan |last6=Matena |first6=Michael |last7=Zhou |first7=Yanqi |last8=Li |first8=Wei |last9=Liu |first9=Peter J. |date=2020 |title=Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer |url=http://jmlr.org/papers/v21/20-074.html |journal=Journal of Machine Learning Research |volume=21 |issue=140 |pages=1–67 |arxiv=1910.10683 |issn=1533-7928}}</ref><ref name=":6">{{Citation |title=google-research/text-to-text-transfer-transformer |date=2024-08-21 |url=https://github.com/google-research/text-to-text-transfer-transformer |access-date=2024-08-21 |publisher=Google Research}}</ref> Like the original Transformer model,<ref>{{Cite journal |last1=Vaswani |first1=Ashish |last2=Shazeer |first2=Noam |last3=Parmar |first3=Niki |last4=Uszkoreit |first4=Jakob |last5=Jones |first5=Llion |last6=Gomez |first6=Aidan N |last7=Kaiser |first7=Łukasz |last8=Polosukhin |first8=Illia |date=2017 |title=Attention is All you Need |url=https://proceedings.neurips.cc/paper_files/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html |journal=Advances in Neural Information Processing Systems |publisher=Curran Associates, Inc. |volume=30}}</ref> T5 models are [[Transformer (deep learning architecture)|encoder-decoder Transformers]], where the encoder processes the input text, and the decoder generates the output text.

T5 models are usually pretrained on a massive dataset of text and code, after which they can perform the text-based tasks that are similar to their pretrained tasks. They can also be finetuned to perform other tasks.

T5 models have been employed in various applications, including chatbots, machine translation systems, text summarization tools, code generation, and robotics.<ref>{{Cite arXiv |eprint=2210.03094 |class=cs.RO |first1=Yunfan |last1=Jiang |first2=Agrim |last2=Gupta |title=VIMA: General Robot Manipulation with Multimodal Prompts |date=2022-10-06 |language=en |last3=Zhang |first3=Zichen |last4=Wang |first4=Guanzhi |last5=Dou |first5=Yongqiang |last6=Chen |first6=Yanjun |last7=Fei-Fei |first7=Li |last8=Anandkumar |first8=Anima |last9=Zhu |first9=Yuke}}</ref>

== Training ==
The original T5 models are pre-trained on the [[Colossal Clean Crawled Corpus]] (C4), containing text and code [[Web crawler|scraped from the internet]]. This pre-training process enables the models to learn general language understanding and generation abilities. T5 models can then be fine-tuned on specific downstream tasks, adapting their knowledge to perform well in various applications.

The T5 models were pretrained on many tasks, all in the format of <code><input text></code> -> <code><output text></code>.
[[File:T5-finetune-summarization.svg|thumb|How a T5 can be finetuned for a summarization task.<ref name=":7">{{Cite book |last1=Zhang |first1=Aston |title=Dive into deep learning |last2=Lipton |first2=Zachary |last3=Li |first3=Mu |last4=Smola |first4=Alexander J. |date=2024 |publisher=Cambridge University Press |isbn=978-1-009-38943-3 |location=Cambridge New York Port Melbourne New Delhi Singapore |chapter=11.9. Large-Scale Pretraining with Transformers |chapter-url=https://d2l.ai/chapter_attention-mechanisms-and-transformers/large-pretraining-transformers.html}}</ref>]]
Some examples are:

* restoring corrupted text: <code>Thank you <X> me to your party <Y> week.</code> -> <code><X> for inviting <Y> last <Z></code>, where the <code><Z></code> means "end of output", and the  <code><X></code> and  <code><Y></code> denote blanks to be filled, called "sentinels" in the original report.
* translation: <code>translate English to German: That is good.</code> -> <code>Das ist gut.</code>.
* judging the grammatical acceptability of a sentence ([[Corpus of Linguistic Acceptability|CoLA]] sentence): <code>The course is jumping well.</code> -> <code>not acceptable</code> .

== Architecture ==
[[File:T5_encoder-decoder_structure.svg|thumb|T5 encoder-decoder structure, showing the attention structure. In the encoder self-attention (lower square), all input tokens attend to each other; In the encoder–decoder cross-attention (upper rectangle), each target token attends to all input tokens; In the decoder self-attention (upper triangle), each target token attends to present and past target tokens only (causal).<ref name=":7" />]]
The T5 series encompasses several models with varying sizes and capabilities, all [[Transformer (deep learning architecture)|encoder-decoder Transformers]], where the encoder processes the input text, and the decoder generates the output text.

These models are often distinguished by their parameter count, which indicates the complexity and potential capacity of the model. The original paper<ref name=":0" /> reported the following 5 models:

{| class="wikitable sortable"
|+T5 properties{{NoteTag|text=<syntaxhighlight lang="python">
import torch
from transformers import AutoConfig, AutoModelForSeq2SeqLM

def count_parameters(model):
    enc = sum(p.numel() for p in model.encoder.parameters())
    dec = sum(p.numel() for p in model.decoder.parameters())
    total = enc + dec
    return total, enc, dec

for name in ["t5-small", "t5-base", "t5-large", "t5-3b", "t5-11b"]:
    print(f"Model: {name}")
    config = AutoConfig.from_pretrained(f"google-t5/{name}")
    torch_dtype = torch.float16

    model = AutoModelForSeq2SeqLM.from_config(config,
                                              torch_dtype=torch_dtype)

    total, enc, dec = count_parameters(model)

    print(f"Total number of parameters in {name}: {total}")
    print(f"Total number of parameters in encoder: {enc}")
    print(f"Total number of parameters in decoder: {dec}")

    del model
</syntaxhighlight>}}
|-
! Name !! Total parameters 
!Encoder parameters
!Decoder parameters!! <math>n_{\text{layer}}</math>!! <math>d_{\text{model}}</math> !! <math>d_{\text{ff}}</math> !! <math>d_{\text{kv}}</math> !! <math>n_{\text{head}}</math>
|-
| Small || 76,956,160 
|35,330,816
|41,625,344|| 6 || 512 || 2048 || 64 || 8
|-
| Base || 247,577,856 
|109,628,544
|137,949,312|| 12 || 768 || 3072 || 64 || 12
|-
| Large || 770,567,168 
|334,939,648
|435,627,520|| 24 || 1024 || 4096 || 64 || 16
|-
| 3B || 2,884,497,408 
|1,240,909,824
|1,643,587,584|| 24 || 1024 || 16384 || 128 || 32
|-
| 11B ||11,340,220,416
|4,864,791,552
| 6,475,428,864|| 24 || 1024 || 65536 || 128 || 128
|}
<sup>*The encoder and the decoder have the same shape. So for example, the T5-small has 6 layers in the encoder and 6 layers in the decoder.</sup>

In the above table,

* <math>n_{\text{layer}}</math>: Number of layers in the encoder; also, number of layers in the decoder. They always have the same number of layers.
* <math>n_{\text{head}}</math>: Number of attention heads in each attention block.
* <math>d_{\text{model}}</math>: Dimension of the embedding vectors.
* <math>d_{\text{ff}}</math>: Dimension of the feedforward network within each encoder and decoder layer.
* '''<math>d_{\text{kv}}</math>''': Dimension of the key and value vectors used in the self-attention mechanism.
Note that unlike typical Transformers, the 3B and 11B models do not satisfy <math>d_{\text{model}} = d_{\text{kv}} n_{\text{head}}</math>.<ref>{{Cite web |date=2020-04-24 |title=config.json · google-t5/t5-11b at main |url=https://huggingface.co/google-t5/t5-11b/blob/main/config.json |access-date=2024-09-17 |website=huggingface.co}}</ref>

Compared to the original Transformer, it uses a few minor modifications: layer normalization with no additive bias; placing the layer normalization outside the residual path; relative positional embedding.<ref>{{Citation |last1=Shaw |first1=Peter |title=Self-Attention with Relative Position Representations |date=2018-04-12 |arxiv=1803.02155 |last2=Uszkoreit |first2=Jakob |last3=Vaswani |first3=Ashish}}</ref>

For all experiments, they used a WordPiece tokenizer, with vocabulary size 32,000. The tokenizer is shared across both the input and output of each model. It was trained on a mixture of [[English language|English]], [[German language|German]], [[French language|French]], and [[Romanian language|Romanian]] data from the C4 dataset, at a ratio of 10:1:1:1.

== Variants ==
Several subsequent models used the T5 architecture, with non-standardized naming conventions used to differentiate them. This section attempts to collect the main ones. An exhaustive list of the variants released by Google Brain is on the GitHub repo for T5X.<ref name=":5">{{Cite web |title=t5x/docs/models.md at main · google-research/t5x |url=https://github.com/google-research/t5x/blob/main/docs/models.md |access-date=2024-08-05 |website=GitHub |language=en}}</ref>

Some models are trained from scratch while others are trained by starting with a previous trained model. By default, each model is trained from scratch, except otherwise noted.

* ''T5'' small, base, large, 3B, 11B (2019): The original models.<ref name=":0" /> 
* ''T5 1.1'' small, base, large, XL, XXL: Improved versions of the original T5 series. These have roughly equal parameters. The [[activation function]] is GEGLU<ref>{{Citation |last=Shazeer |first=Noam |title=GLU Variants Improve Transformer |date=2020-02-12 |arxiv=2002.05202}}</ref> instead of ReLU. The 3B and the 11B were changed to "XL" and "XXL", and their shapes are changed:<ref name=":5" /><ref>{{Cite web |date=2020-11-19 |title=config.json · google/t5-v1_1-xl at main |url=https://huggingface.co/google/t5-v1_1-xl/blob/main/config.json |access-date=2024-09-17 |website=huggingface.co}}</ref><ref>{{Cite web |date=2020-11-19 |title=config.json · google/t5-v1_1-xxl at main |url=https://huggingface.co/google/t5-v1_1-xxl/blob/main/config.json |access-date=2024-09-17 |website=huggingface.co}}</ref>
{| class="wikitable sortable"
|+T5 v1.1 properties{{NoteTag|text=<syntaxhighlight lang="python">
import torch
from transformers import AutoConfig, AutoModelForSeq2SeqLM

def count_parameters(model):
    enc = sum(p.numel() for p in model.encoder.parameters())
    dec = sum(p.numel() for p in model.decoder.parameters())
    total = enc + dec
    return total, enc, dec

for name in ["small", "base", "large", "xl", "xxl"]:
    print(f"Model: {name}")
    config = AutoConfig.from_pretrained(f"google/t5-v1_1-{name}")
    torch_dtype = torch.float16

    model = AutoModelForSeq2SeqLM.from_config(config,
                                              torch_dtype=torch_dtype)

    total, enc, dec = count_parameters(model)

    print(f"Total number of parameters in {name}: {total}")
    print(f"Total number of parameters in encoder: {enc}")
    print(f"Total number of parameters in decoder: {dec}")

    del model
</syntaxhighlight>}}
|-
! Name !! Total parameters 
!Encoder parameters
!Decoder parameters!! <math>n_{\text{layer}}</math> !! <math>d_{\text{model}}</math> !! <math>d_{\text{ff}}</math> !! <math>d_{\text{kv}}</math> !! <math>n_{\text{head}}</math>
|-
| Small || 76,961,152 
|35,332,800
|41,628,352|| 8 || 512 || 1024 || 64 || 6
|-
| Base || 247,577,856 
|109,628,544
|137,949,312|| 12 || 768 || 2048 || 64 || 12
|-
| Large || 783,150,080 
|341,231,104
|441,918,976|| 24 || 1024 || 2816 || 64 || 16
|-
| XL || 2,849,757,184 
|1,223,527,424
|1,626,229,760|| 24 || 2048 || 5120 || 64 || 32
|-
| XXL || 11,135,332,352 
|4,762,310,656
| 6,373,021,696|| 24 || 4096 || 10240 || 64 || 64
|}

* ''LM-adapted T5'' (2021): a series of models (from small to XXL) that started from checkpoints of the ''T5'' series, but trained further on 100B additional tokens from C4.<ref>{{Citation |last1=Lester |first1=Brian |title=The Power of Scale for Parameter-Efficient Prompt Tuning |date=2021-09-02 |arxiv=2104.08691 |last2=Al-Rfou |first2=Rami |last3=Constant |first3=Noah}}</ref>
* Switch Transformer (2021): a [[Mixture of experts|mixture-of-experts]] variant of T5, by replacing the feedforward layers in the encoder and decoder blocks with mixture of expert feedforward layers.<ref>{{Citation |last1=Fedus |first1=William |title=Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity |date=2022-06-16 |arxiv=2101.03961 |last2=Zoph |first2=Barret |last3=Shazeer |first3=Noam}}</ref><ref>{{Cite web |title=SwitchTransformers |url=https://huggingface.co/docs/transformers/en/model_doc/switch_transformers |access-date=2024-08-05 |website=huggingface.co}}</ref>
* ''T0'' 3B, 11B (2021): a series of models that started from checkpoints of ''LM-adapted T5'', and further trained to perform tasks based only on task instruction ([[Zero-shot learning|zero-shot]]).<ref>{{Citation |last1=Sanh |first1=Victor |title=Multitask Prompted Training Enables Zero-Shot Task Generalization |date=2022-03-17 |arxiv=2110.08207 |last2=Webson |first2=Albert |last3=Raffel |first3=Colin |last4=Bach |first4=Stephen H. |last5=Sutawika |first5=Lintang |last6=Alyafeai |first6=Zaid |last7=Chaffin |first7=Antoine |last8=Stiegler |first8=Arnaud |last9=Scao |first9=Teven Le}}</ref> Different entries in the series uses different finetuning data.<ref>{{Cite web |date=2024-03-04 |title=bigscience/T0 · Hugging Face |url=https://huggingface.co/bigscience/T0 |access-date=2024-08-21 |website=huggingface.co}}</ref>
* ''ByT5'' (2021): a byte-level version of T5, trained on mC4 (multilingual C4) dataset.<ref>{{Cite journal |last1=Xue |first1=Linting |last2=Barua |first2=Aditya |last3=Constant |first3=Noah |last4=Al-Rfou |first4=Rami |last5=Narang |first5=Sharan |last6=Kale |first6=Mihir |last7=Roberts |first7=Adam |last8=Raffel |first8=Colin |date=2022-03-25 |title=ByT5: Towards a Token-Free Future with Pre-trained Byte-to-Byte Models |url=https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00461/110049/ByT5-Towards-a-Token-Free-Future-with-Pre-trained |journal=Transactions of the Association for Computational Linguistics |language=en |volume=10 |pages=291–306 |doi=10.1162/tacl_a_00461 |issn=2307-387X|arxiv=2105.13626 }}</ref> It operates on text encoded as [[UTF-8]] bytes, without tokenizers.
* ''Flan-T5-XL'' (2022): a model that started with a checkpoint of ''T5 XL'', then [[Instruction tuning|instruction-tuned]] on the FLAN dataset.<ref>{{Cite journal |last1=Chung |first1=Hyung Won |last2=Hou |first2=Le |last3=Longpre |first3=Shayne |last4=Zoph |first4=Barret |last5=Tay |first5=Yi |last6=Fedus |first6=William |last7=Li |first7=Yunxuan |last8=Wang |first8=Xuezhi |last9=Dehghani |first9=Mostafa |last10=Brahma |first10=Siddhartha |last11=Webson |first11=Albert |last12=Gu |first12=Shixiang Shane |last13=Dai |first13=Zhuyun |last14=Suzgun |first14=Mirac |last15=Chen |first15=Xinyun |date=2024 |title=Scaling Instruction-Finetuned Language Models |url=http://jmlr.org/papers/v25/23-0870.html |journal=Journal of Machine Learning Research |volume=25 |issue=70 |pages=1–53 |arxiv=2210.11416 |issn=1533-7928}}</ref><ref>{{Cite journal |last1=Longpre |first1=Shayne |last2=Hou |first2=Le |last3=Vu |first3=Tu |last4=Webson |first4=Albert |last5=Chung |first5=Hyung Won |last6=Tay |first6=Yi |last7=Zhou |first7=Denny |last8=Le |first8=Quoc V. |last9=Zoph |first9=Barret |last10=Wei |first10=Jason |last11=Roberts |first11=Adam |date=2023-07-03 |title=The Flan Collection: Designing Data and Methods for Effective Instruction Tuning |url=https://proceedings.mlr.press/v202/longpre23a.html |journal=Proceedings of the 40th International Conference on Machine Learning |language=en |publisher=PMLR |pages=22631–22648|arxiv=2301.13688 }}</ref><ref name=":2">{{Citation |title=google-research/FLAN |date=2024-08-03 |url=https://github.com/google-research/FLAN |access-date=2024-08-05 |publisher=Google Research}}</ref><ref>{{Cite web |date=2024-01-04 |title=google/flan-t5-xl · Hugging Face |url=https://huggingface.co/google/flan-t5-xl |access-date=2024-08-05 |website=huggingface.co}}</ref>
* ''T5X'' (2022): a [[Google JAX|JAX]]-based re-implementation of the original ''T5'' codebase. It is '''not''' a model.<ref name=":1">{{Cite journal |last1=Roberts |first1=Adam |last2=Chung |first2=Hyung Won |last3=Mishra |first3=Gaurav |last4=Levskaya |first4=Anselm |last5=Bradbury |first5=James |last6=Andor |first6=Daniel |last7=Narang |first7=Sharan |last8=Lester |first8=Brian |last9=Gaffney |first9=Colin |last10=Mohiuddin |first10=Afroz |last11=Hawthorne |first11=Curtis |last12=Lewkowycz |first12=Aitor |last13=Salcianu |first13=Alex |last14=Zee |first14=Marc van |last15=Austin |first15=Jacob |date=2023 |title=Scaling Up Models and Data with t5x and seqio |url=http://jmlr.org/papers/v24/23-0795.html |journal=Journal of Machine Learning Research |volume=24 |issue=377 |pages=1–8 |issn=1533-7928}}</ref> The original T5 codebase was implemented in [[TensorFlow]] with MeshTF.<ref name=":6" />
* ''UL2'' 20B (2022): a model with the same architecture as the ''T5'' series, but scaled up to 20B, and trained with "mixture of denoisers" objective on the C4.<ref name=":3">{{Citation |last1=Tay |first1=Yi |title=UL2: Unifying Language Learning Paradigms |date=2023-02-28 |arxiv=2205.05131 |last2=Dehghani |first2=Mostafa |last3=Tran |first3=Vinh Q. |last4=Garcia |first4=Xavier |last5=Wei |first5=Jason |last6=Wang |first6=Xuezhi |last7=Chung |first7=Hyung Won |last8=Shakeri |first8=Siamak |last9=Bahri |first9=Dara}}</ref> It was trained on a TPU cluster by accident, when a training run was left running accidentally for a month.<ref>{{Cite web |title=Training great LLMs entirely from ground up in the wilderness as a startup |url=https://www.yitay.net/blog/training-great-llms-entirely-from-ground-zero-in-the-wilderness |access-date=2024-10-18 |website=Yi Tay |language=en-US}}</ref>
* ''Flan-UL2'' 20B (2022): ''UL2'' 20B [[Instruction tuning|instruction-finetuned]] on the FLAN dataset.<ref name=":3" /><ref name=":2" />
* ''Pile-T5'' (2024): has the same architecture of ''T5'', except it used the [[Llama (language model)|Llama]] tokenizer. It was trained on [[The Pile (dataset)|The Pile]]. It came in sizes of base, large, XL, XXL.<ref name=":4">{{Cite web |last1=Sutawika |first1=Lintang |last2=Komatsuzaki |first2=Aran |last3=Raffel |first3=Colin |date=2024-04-15 |title=Pile-T5 |url=https://blog.eleuther.ai/pile-t5/ |access-date=2024-05-05 |website=EleutherAI Blog |language=en}}</ref>

== Applications ==
The T5 model itself is an encoder-decoder model, allowing it to be used for instruction following. The encoder encodes the instruction, and the decoder autoregressively generates the reply.

The T5 encoder can be used as a text encoder, much like BERT. It encodes a text into a sequence of real-number vectors, which can be used for downstream applications. For example, Google Imagen<ref>{{Cite web |title=Imagen: Text-to-Image Diffusion Models |url=https://imagen.research.google/ |access-date=2024-08-23 |website=imagen.research.google}}</ref> uses ''T5-XXL'' as text encoder, and the encoded text vectors are used as conditioning on a [[diffusion model]]. As another example, the AuraFlow diffusion model<ref>{{Cite web |title=AuraFlow |url=https://huggingface.co/docs/diffusers/main/en/api/pipelines/aura_flow |access-date=2024-08-23 |website=huggingface.co}}</ref> uses ''Pile-T5-XL''.

== References ==
<references />

== External links ==

* {{Cite web |date=2024-07-31 |title=T5 release - a google Collection |url=https://huggingface.co/collections/google/t5-release-65005e7c520f8d7b4d037918 |access-date=2024-10-16 |website=huggingface.co}}

== Notes ==
<references group="note" responsive="1"></references>
{{Google AI}}
{{Google LLC}}
{{Natural Language Processing}}
{{Artificial intelligence navbox}}

[[Category:Google software]]
[[Category:Large language models]]
[[Category:Software using the Apache license]]
[[Category:Open-source artificial intelligence]]