Edit model card

Kreyòl-MT

Welcome to the repository for our from-scratch public-data model.

Please see our paper: 📄 "Kreyòl-MT: Building Machine Translation for Latin American, Caribbean, and Colonial African Creole Languages"

And our GitHub repository: 💻 Kreyòl-MT

And cite our work:

@article{robinson2024krey,
  title={Krey$\backslash$ol-MT: Building MT for Latin American, Caribbean and Colonial African Creole Languages},
  author={Robinson, Nathaniel R and Dabre, Raj and Shurtz, Ammon and Dent, Rasul and Onesi, Onenamiyi and Monroc, Claire Bizon and Grobol, Lo{\"\i}c and Muhammad, Hasan and Garg, Ashi and Etori, Naome A and others},
  journal={arXiv preprint arXiv:2405.05376},
  year={2024}
}

Model hosted here

This is a many-to-many model for translation into and out of Creole languages, trained from scratch on public data.

from transformers import MBartForConditionalGeneration, AutoModelForSeq2SeqLM
from transformers import AlbertTokenizer, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("jhu-clsp/kreyol-mt-scratch-pubtrain", do_lower_case=False, use_fast=False, keep_accents=True)

# The tokenizer we use is based on the AlbertTokenizer class which is essentially sentencepiece. We train this sentencepeice model from scratch.
# Or use tokenizer = AlbertTokenizer.from_pretrained("jhu-clsp/kreyol-mt-scratch-pubtrain", do_lower_case=False, use_fast=False, keep_accents=True)

model = AutoModelForSeq2SeqLM.from_pretrained("jhu-clsp/kreyol-mt-scratch-pubtrain")

# Or use model = MBartForConditionalGeneration.from_pretrained("jhu-clsp/kreyol-mt-scratch-pubtrain")

# Some initial mapping
bos_id = tokenizer._convert_token_to_id_with_added_voc("<s>")
eos_id = tokenizer._convert_token_to_id_with_added_voc("</s>")
pad_id = tokenizer._convert_token_to_id_with_added_voc("<pad>")

# First tokenize the input and outputs. The format below is how the model was trained so the input should be "Sentence </s> <2hwc>". Similarly, the output should be "<2eng> Sentence </s>".
# Example: For Saint Lucian Patois to English translation, we need to use language indicator tags: <2acf> and <2eng> where acf represents Saint Lucian Patois and eng represents English.
# The following language indicator tokens are usable: <2acf>, <2aoa>, <2ara>, <2aze>, <2bah>, <2brc>, <2bzj>, <2bzk>, <2cab>, <2ceb>, <2cri>, <2crs>, <2dcr>, <2deu>, <2djk>, <2eng>, <2fab>, <2fng>, <2fpe>, <2fra>, <2gcf>, <2gcr>, <2gpe>, <2gul>, <2gyn>, <2hat>, <2icr>, <2jam>, <2kea>, <2kri>, <2ktu>, <2lou>, <2mart1259>, <2mfe>, <2mue>, <2nep>, <2pap>, <2pcm>, <2por>, <2pov>, <2pre>, <2rcf>, <2sag>, <2spa>, <2srm>, <2srn>, <2svc>, <2tpi>, <2trf>, <2wes>, <2zho>
# For what language each language code corresponds to please look here: https://github.com/JHU-CLSP/Kreyol-MT?tab=readme-ov-file#building-machine-translation-for-latin-american-caribbean-and-colonial-african-creole-languages

inp = tokenizer('Mi tingk se yu de tel mi lai. </s> <2jam>', add_special_tokens=False, return_tensors="pt", padding=True).input_ids

model.eval() # Set dropouts to zero

model_output=model.generate(inp, use_cache=True, num_beams=4, max_length=60, min_length=1, early_stopping=True, pad_token_id=pad_id, bos_token_id=bos_id, eos_token_id=eos_id, decoder_start_token_id=tokenizer._convert_token_to_id_with_added_voc("<2eng>"))

decoded_output=tokenizer.decode(model_output[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)

print(decoded_output)

results

Downloads last month
13
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.