Hugging Face's logo
language:
- om
- am
- rw
- rn
- ha
- ig
- pcm
- so
- sw
- ti
- yo
- multilingual
datasets:
AfriBERTa_small
Model description
AfriBERTa small is a pretrained multilingual language model with around 97 million parameters. The model has 4 layers, 6 attention heads, 768 hidden units and 3072 feed forward size. The model was pretrained on 11 African languages namely - Afaan Oromoo (also called Oromo), Amharic, Gahuza (a mixed language containing Kinyarwanda and Kirundi), Hausa, Igbo, Nigerian Pidgin, Somali, Swahili, Tigrinya and Yorùbá. The model has been shown to obtain competitive downstream performances on text classification and Named Entity Recognition on several African languages, including those it was not pretrained on.
Intended uses & limitations
How to use
You can use this model with Transformers for any downstream task. For example, assuming we want to finetune this model on a token classification task, we do the following:
>>> from transformers import AutoTokenizer, AutoModelForTokenClassification
>>> model = AutoModelForTokenClassification.from_pretrained("castorini/afriberta_small")
>>> tokenizer = AutoTokenizer.from_pretrained("castorini/afriberta_small")
# we have to manually set the model max length because it is an imported sentencepiece model which hugginface does not properly support right now
>>> tokenizer.model_max_length = 512
Limitations and bias
This model is possibly limited by its training dataset which are majorly obtained from news articles from a specific span of time. Thus, it may not generalize well.
Training data
The model was trained on an aggregation of datasets from the BBC news website and Common Crawl.
Training procedure
For information on training procedures, please refer to the AfriBERTa paper or repository
BibTeX entry and citation info
Kelechi Ogueji, Yuxin Zhu, Jimmy Lin.
Small Data? No Problem! Exploring the Viability of Pretrained Multilingual Language Models for Low-resourced Languages
Proceedings of the 1st workshop on Multilingual Representation Learning at EMNLP 2021