Edit model card

Model Description

This is the model presented in the paper "Exploring Methods for Cross-lingual Text Style Transfer: The Case of Text Detoxification".

The model is based on mBART-large-50 and trained on two parallel detoxification corpora: ParaDetox and RuDetox. More details about this model are in the paper.

Usage

  1. Model loading.
from transformers import MBartForConditionalGeneration, AutoTokenizer

model = MBartForConditionalGeneration.from_pretrained("s-nlp/mbart-detox-en-ru").cuda()
tokenizer = AutoTokenizer.from_pretrained("facebook/mbart-large-50")
  1. Detoxification utility.
def paraphrase(text, model, tokenizer, n=None, max_length="auto", beams=3):
    texts = [text] if isinstance(text, str) else text
    inputs = tokenizer(texts, return_tensors="pt", padding=True)["input_ids"].to(
        model.device
    )
    if max_length == "auto":
        max_length = inputs.shape[1] + 10

    result = model.generate(
        inputs,
        num_return_sequences=n or 1,
        do_sample=True,
        temperature=1.0,
        repetition_penalty=10.0,
        max_length=max_length,
        min_length=int(0.5 * max_length),
        num_beams=beams,
        forced_bos_token_id=tokenizer.lang_code_to_id[tokenizer.tgt_lang]
    )
    texts = [tokenizer.decode(r, skip_special_tokens=True) for r in result]

    if not n and isinstance(text, str):
        return texts[0]
    return texts

Citation

TBD

Downloads last month
63
Safetensors
Model size
611M params
Tensor type
F32
·

Datasets used to train s-nlp/mbart-detox-en-ru