metadata
license: mit
This is a many-to-many model for Creole-English, English-Creole and Creole-Creole MT, fine-tuned on top of facebook/mbart-large-50-many-to-many-mmt, with only public data.
Usage:
from transformers import MBartForConditionalGeneration, AutoModelForSeq2SeqLM
from transformers import MbartTokenizer, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("n8rob/kreyol-mt-pubtrain", do_lower_case=False, use_fast=False, keep_accents=True)
# Or use tokenizer = MbartTokenizer.from_pretrained("n8rob/kreyol-mt-pubtrain", use_fast=False)
model = AutoModelForSeq2SeqLM.from_pretrained("n8rob/kreyol-mt-pubtrain")
# Or use model = MBartForConditionalGeneration.from_pretrained("n8rob/kreyol-mt-pubtrain")
# First tokenize the input and outputs. The format below is how the model was trained so the input should be "Sentence </s> SRCCODE". Similarly, the output should be "TGTCODE Sentence </s>".
# For Hawaiian Creole to English translation, we need to use language indicator tags: fr_XX and en_XX where fr_XX represents Hawaiian Creole (hwc) and en_XX represents English (eng).
# For a mapping of the original language and language code (3 character) to mBART-50 compatible language tokens consider the following dictionary:
# {} ## Add this
# Note: We mapped languages to their language tokens manually. For example, we used en_XX, fr_XX, es_XX for English, French and Spanish as in the original mBART-50 model. But then we repurposed other tokens for Creoles.
# As for what the language codes and their corresponding languages are, please refer to: 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('Wen dey wen stretch him out fo whip him real hard , Paul wen tell da captain dat stay dea , "Dis okay in da rules fo da Rome peopo ? fo you fo whip one guy dat get da same rights jalike da Rome peopo ? even one guy dat neva do notting wrong ?" </s> fr_XX', 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("en_XX"))
decoded_output=tokenizer.decode(model_output[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)
print(decoded_output)