voxreality/src_ctx_and_term_nllb_600M

The model and the tokenizer are based on facebook/nllb-200-distilled-600M.

We trained the model to use one sentence of context and a single term of the terminology-constraint. The context is prepended to the input sentence with the sep_token in between. The term should be in the target language and be postpended to the input sentence with the sep_token in between. In case of no terminology constraint, the sep_token should also be added. We used a subset of the OpenSubtitles2018 dataset for training. We trained on the interleaved dataset for all directions between the following languages: English, German, Dutch, Spanish, Italian, and Greek.

The tokenizer of the base model was not changed. For the language codes, see the base model.

Use this code for translation:

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_name = 'voxreality/src_ctx_and_term_nllb_600M'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

max_length = 100
src_lang = 'eng_Latn'
tgt_lang = 'deu_Latn'
context_text = 'This is an optional context sentence.'
target_term = 'text'  # term to be used in the target language
sentence_text = 'Text to be translated.'

# if a context and a term are provided use the following:
input_text = f'{context_text} {tokenizer.sep_token} {sentence_text} {tokenizer.sep_token} {target_term}'
# if no context but a term is provided use the following:
# input_text = f'{sentence_text} {tokenizer.sep_token} {target_term}'
# if a context is provided but no term use the following:
# input_text = f'{context_text} {tokenizer.sep_token} {sentence_text} {tokenizer.sep_token}'
# if not context nor term is provided use the following:
# input_text = f'{sentence_text} {tokenizer.sep_token}'

tokenizer.src_lang = src_lang
inputs = tokenizer(input_text, return_tensors='pt').to(model.device)
model_output = model.generate(**inputs,
                              forced_bos_token_id=tokenizer.lang_code_to_id[tgt_lang],
                              max_length=max_length)
output_text = tokenizer.batch_decode(model_output, skip_special_tokens=True)[0]

print(output_text)

You can also use the pipeline

from transformers import pipeline

model_name = 'voxreality/src_ctx_and_term_nllb_600M'
translation_pipeline = pipeline("translation", model=model_name)
sep_token = translation_pipeline.tokenizer.sep_token
src_lang = 'eng_Latn'
tgt_lang = 'deu_Latn'
context_text = 'This is an optional context sentence.'
target_term = 'text'  # term to be used in the target language
sentence_text = 'Text to be translated.'


# if a context and a term are provided use the following:
input_texts = [f'{context_text} {sep_token} {sentence_text} {sep_token} {target_term}']
# if no context but a term is provided use the following:
# input_texts = [f'{sentence_text} {sep_token} {target_term}']
# if a context is provided but no term use the following:
# input_texts = [f'{context_text} {sep_token} {sentence_text} {sep_token}']
# if not context nor term is provided use the following:
# input_texts = [f'{sentence_text} {sep_token}']

pipeline_output = translation_pipeline(input_texts, src_lang=src_lang, tgt_lang=tgt_lang)

print(pipeline_output[0]['translation_text'])