---
license: apache-2.0
language:
- en
- ru
- es
- fr
- de
- it
- pt
- pl
- nl
- vi
- tr
- sv
- id
- ro
- cs
- zh
- hu
- ja
- th
- fi
- fa
- uk
- da
- el
- "no"
- bg
- sk
- ko
- ar
- lt
- ca
- sl
- he
- et
- lv
- hi
- sq
- ms
- az
- sr
- ta
- hr
- kk
- is
- ml
- mr
- te
- af
- gl
- fil
- be
- mk
- eu
- bn
- ka
- mn
- bs
- uz
- ur
- sw
- yue
- ne
- kn
- kaa
- gu
- si
- cy
- eo
- la
- hy
- ky
- tg
- ga
- mt
- my
- km
- tt
- so
- ku
- ps
- pa
- rw
- lo
- ha
- dv
- fy
- lb
- ckb
- mg
- gd
- am
- ug
- ht
- grc
- hmn
- sd
- jv
- mi
- tk
- ceb
- yi
- ba
- fo
- or
- xh
- su
- kl
- ny
- sm
- sn
- co
- zu
- ig
- yo
- pap
- st
- haw
- as
- oc
- cv
- lus
- tet
- gsw
- sah
- br
- rm
- sa
- bo
- om
- se
- ce
- cnh
- ilo
- hil
- udm
- os
- lg
- ti
- vec
- ts
- tyv
- kbd
- ee
- iba
- av
- kha
- to
- tn
- nso
- fj
- zza
- ak
- ada
- otq
- dz
- bua
- cfm
- ln
- chm
- gn
- krc
- wa
- hif
- yua
- srn
- war
- rom
- bik
- pam
- sg
- lu
- ady
- kbp
- syr
- ltg
- myv
- iso
- kac
- bho
- ay
- kum
- qu
- za
- pag
- ngu
- ve
- pck
- zap
- tyz
- hui
- bbc
- tzo
- tiv
- ksd
- gom
- min
- ang
- nhe
- bgp
- nzi
- nnb
- nv
- zxx
- bci
- kv
- new
- mps
- alt
- meu
- bew
- fon
- iu
- abt
- mgh
- mnw
- tvl
- dov
- tlh
- ho
- kw
- mrj
- meo
- crh
- mbt
- emp
- ace
- ium
- mam
- gym
- mai
- crs
- pon
- ubu
- fip
- quc
- gv
- kj
- btx
- ape
- chk
- rcf
- shn
- tzh
- mdf
- ppk
- ss
- gag
- cab
- kri
- seh
- ibb
- tbz
- bru
- enq
- ach
- cuk
- kmb
- wo
- kek
- qub
- tab
- bts
- kos
- rwo
- cak
- tuc
- bum
- cjk
- gil
- stq
- tsg
- quh
- mak
- arn
- ban
- jiv
- sja
- yap
- tcy
- toj
- twu
- xal
- amu
- rmc
- hus
- nia
- kjh
- bm
- guh
- mas
- acf
- dtp
- ksw
- bzj
- din
- zne
- mad
- msi
- mag
- mkn
- kg
- lhu
- ch
- qvi
- mh
- djk
- sus
- mfe
- srm
- dyu
- ctu
- gui
- pau
- inb
- bi
- mni
- guc
- jam
- wal
- jac
- bas
- gor
- skr
- nyu
- noa
- sda
- gub
- nog
- cni
- teo
- tdx
- sxn
- rki
- nr
- frp
- alz
- taj
- lrc
- cce
- rn
- jvn
- hvn
- nij
- dwr
- izz
- msm
- bus
- ktu
- chr
- maz
- tzj
- suz
- knj
- bim
- gvl
- bqc
- tca
- pis
- prk
- laj
- mel
- qxr
- niq
- ahk
- shp
- hne
- spp
- koi
- krj
- quf
- luz
- agr
- tsc
- mqy
- gof
- gbm
- miq
- dje
- awa
- bjj
- qvz
- sjp
- tll
- raj
- kjg
- bgz
- quy
- cbk
- akb
- oj
- ify
- mey
- ks
- cac
- brx
- qup
- syl
- jax
- ff
- ber
- tks
- trp
- mrw
- adh
- smt
- srr
- ffm
- qvc
- mtr
- ann
- kaa
- aa
- noe
- nut
- gyn
- kwi
- xmm
- msb
library_name: transformers
tags:
- text-generation-inference
datasets:
- allenai/MADLAD-400
pipeline_tag: translation
---

T5ForConditionalGeneration files for Google's [Madlad-400](https://github.com/google-research/google-research/tree/master/madlad_400) 3B parameter MT model.

Article: [MADLAD-400: A Multilingual And Document-Level Large Audited Dataset](https://arxiv.org/abs/2309.04662)

Available models:
- [3B](https://huggingface.co/jbochi/madlad400-3b-mt)
- [7B](https://huggingface.co/jbochi/madlad400-7b-mt)
- [7B-BT](https://huggingface.co/jbochi/madlad400-7b-mt-bt)
- [10B](https://huggingface.co/jbochi/madlad400-10b-mt)

Abstract:

> We introduce MADLAD-400, a manually audited, general domain 3T token monolingual dataset based on CommonCrawl, spanning 419 languages. We discuss the limitations revealed by self-auditing MADLAD-400, and the role data auditing had in the dataset creation process. We then train and release a 10.7B-parameter multilingual machine translation model on 250 billion tokens covering over 450 languages using publicly available data, and find that it is competitive with models that are significantly larger, and report the results on different domains. In addition, we train a 8B-parameter language model, and assess the results on few-shot translation. We make the baseline models available to the research community.

## Usage

Usage with Huggingface's transformers:

```python
from transformers import T5ForConditionalGeneration, T5Tokenizer, GenerationConfig

model = T5ForConditionalGeneration.from_pretrained('jbochi/madlad400-3b-mt')
tokenizer = T5Tokenizer.from_pretrained('jbochi/madlad400-3b-mt')

text = "<2pt> I love pizza!"
input_ids = tokenizer(text, return_tensors="pt").input_ids
outputs = model.generate(input_ids=input_ids)

tokenizer.decode(outputs[0], skip_special_tokens=True)
# Eu adoro pizza!
```

Usage with [candle](https://github.com/huggingface/candle):

```bash
$ cargo run --example t5 --release  -- \
  --model-id "jbochi/madlad400-3b-mt" \
  --prompt "<2de> How are you, my friend?" \
  --decode --temperature 0
```

We also provide a quantized model (1.65 GB vs the original 11.8 GB file):

```
cargo run --example quantized-t5 --release  -- \
  --model-id "jbochi/madlad400-3b-mt" --weight-file "model-q4k.gguf" \
  --prompt "<2de> How are you, my friend?" \
  --temperature 0
...
 Wie geht es dir, mein Freund?
```


## Model conversion

I'm not affiliated with Google and was not involved in this research.

The colab I used to generate these files is [here](https://colab.research.google.com/drive/1rZ2NRyl2zwmg0sQ2Wi-uZZF48iVYulTC#scrollTo=pVODoE6gA9sw).

Quantization was done with candle following this [instruction](https://github.com/huggingface/candle/tree/main/candle-examples/examples/quantized-t5#generating-quantized-weight-files).