Hugging Face's logo

Davlan
/
afro-xlmr-large-75L

Fill-Mask
Transformers
PyTorch
Safetensors
xlm-roberta
Generated from Trainer
Inference Endpoints
Model card Files Community
1
afro-xlmr-large-75L / README.md
Davlan's picture
Davlan
Update README.md
6e178e1
preview code
|
raw
history blame
4.31 kB
metadata
license: mit
tags:
  - generated_from_trainer
model-index:
  - name: afro-xlmr-large-75L
    results: []
language:
  - en
  - am
  - ar
  - so
  - sw
  - pt
  - af
  - fr
  - zu
  - mg
  - ha
  - sn
  - arz
  - ny
  - ig
  - xh
  - yo
  - st
  - rw
  - tn
  - ti
  - ts
  - om
  - run
  - nso
  - ee
  - ln
  - tw
  - pcm
  - gaa
  - loz
  - lg
  - guw
  - bem
  - efi
  - lue
  - lua
  - toi
  - ve
  - tum
  - tll
  - iso
  - kqn
  - zne
  - umb
  - mos
  - tiv
  - lu
  - ff
  - kwy
  - bci
  - rnd
  - luo
  - wal
  - ss
  - lun
  - wo
  - nyk
  - kj
  - ki
  - fon

afro-xlmr-large-75L

AfroXLMR-large was created by MLM adaptation of XLM-R-large model on 75 languages widely spoken in Africa including 4 high resource languages.

Pre-training corpus

A mix of mC4, Wikipedia and OPUS data

Languages

There are 75 languages available :

  • English (eng)
  • Amharic (amh)
  • Arabic (ara)
  • Somali (som)
  • Kiswahili (swa)
  • Portuguese (por)
  • Afrikaans (afr)
  • French (fra)
  • isiZulu (zul)
  • Malagasy (mlg)
  • Hausa (hau)
  • chiShona (sna)
  • Egyptian Arabic (arz)
  • Chichewa (nya)
  • Igbo (ibo)
  • isiXhosa (xho)
  • Yorùbá (yor)
  • Sesotho (sot)
  • Kinyarwanda (kin)
  • Tigrinya (tir)
  • Tsonga (tso)
  • Oromo (orm)
  • Rundi (run)
  • Northern Sotho (nso)
  • Ewe (ewe)
  • Lingala (lin)
  • Twi (twi)
  • Nigerian Pidgin (pcm)
  • Ga (gaa)
  • Lozi (loz)
  • Luganda (lug)
  • Gun (guw)
  • Bemba (bem)
  • Efik (efi)
  • Luvale (lue)
  • Luba-Lulua (lua)
  • Tonga (toi)
  • Tshivenḓa (ven)
  • Tumbuka (tum)
  • Tetela (tll)
  • Isoko (iso)
  • Kaonde (kqn)
  • Zande (zne)
  • Umbundu (umb)
  • Mossi (mos)
  • Tiv (tiv)
  • Luba-Katanga (lub)
  • Fula (fuv)
  • San Salvador Kongo (kwy)
  • Baoulé (bci)
  • Ruund (rnd)
  • Luo (luo)
  • Wolaitta (wal)
  • Swazi (ssw)
  • Lunda (lun)
  • Wolof (wol)
  • Nyaneka (nyk)
  • Kwanyama (kua)
  • Kikuyu (kik)
  • Fon (fon)
  • Bambara (bam)
  • Chokwe (cjk)
  • Dinka (dik)
  • Dyula (dyu)
  • Kabyle (kab)
  • Kamba (kam)
  • Kabiyè (kbp)
  • Kanuri (knc)
  • Kimbundu (kmb)
  • Kikongo (kon)
  • Nuer (nus)
  • Sango (sag)
  • Tamasheq (taq)
  • Tamazight (tzm)

Acknowledgment

We would like to thank Google Cloud for providing us access to TPU v3-8 through the free cloud credits. Model trained using flax, before converted to pytorch.

BibTeX entry and citation info. 

@inproceedings{alabi-etal-2022-adapting,
    title = "Adapting Pre-trained Language Models to {A}frican Languages via Multilingual Adaptive Fine-Tuning",
    author = "Alabi, Jesujoba O.  and
      Adelani, David Ifeoluwa  and
      Mosbach, Marius  and
      Klakow, Dietrich",
    booktitle = "Proceedings of the 29th International Conference on Computational Linguistics",
    month = oct,
    year = "2022",
    address = "Gyeongju, Republic of Korea",
    publisher = "International Committee on Computational Linguistics",
    url = "https://aclanthology.org/2022.coling-1.382",
    pages = "4336--4349",
    abstract = "Multilingual pre-trained language models (PLMs) have demonstrated impressive performance on several downstream tasks for both high-resourced and low-resourced languages. However, there is still a large performance drop for languages unseen during pre-training, especially African languages. One of the most effective approaches to adapt to a new language is language adaptive fine-tuning (LAFT) {---} fine-tuning a multilingual PLM on monolingual texts of a language using the pre-training objective. However, adapting to target language individually takes large disk space and limits the cross-lingual transfer abilities of the resulting models because they have been specialized for a single language. In this paper, we perform multilingual adaptive fine-tuning on 17 most-resourced African languages and three other high-resource languages widely spoken on the African continent to encourage cross-lingual transfer learning. To further specialize the multilingual PLM, we removed vocabulary tokens from the embedding layer that corresponds to non-African writing scripts before MAFT, thus reducing the model size by around 50{\%}. Our evaluation on two multilingual PLMs (AfriBERTa and XLM-R) and three NLP tasks (NER, news topic classification, and sentiment classification) shows that our approach is competitive to applying LAFT on individual languages while requiring significantly less disk space. Additionally, we show that our adapted PLM also improves the zero-shot cross-lingual transfer abilities of parameter efficient fine-tuning methods.",
}