Romanian paraphrase

Fine-tune t5-base model for paraphrase. Since there is no Romanian dataset for paraphrasing, I had to create my own dataset. The dataset contains ~60k examples.

How to use

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("BlackKakapo/t5-base-paraphrase-ro")
model = AutoModelForSeq2SeqLM.from_pretrained("BlackKakapo/t5-base-paraphrase-ro")

Or

from transformers import T5ForConditionalGeneration, T5TokenizerFast 

model = T5ForConditionalGeneration.from_pretrained("BlackKakapo/t5-base-paraphrase-ro")
tokenizer = T5TokenizerFast.from_pretrained("BlackKakapo/t5-base-paraphrase-ro")

Generate

text = "Am impresia că fac multe greșeli."

encoding = tokenizer.encode_plus(text, pad_to_max_length=True, return_tensors="pt")
input_ids, attention_masks = encoding["input_ids"].to(device), encoding["attention_mask"].to(device)

beam_outputs = model.generate(
    input_ids=input_ids, 
    attention_mask=attention_masks,
    do_sample=True,
    max_length=256,
    top_k=10,
    top_p=0.9,
    early_stopping=False,
    num_return_sequences=5
)

for beam_output in beam_outputs:
    text_para = tokenizer.decode(beam_output, skip_special_tokens=True,clean_up_tokenization_spaces=True)
    
    if text.lower() != text_para.lower() or text not in final_outputs:
        final_outputs.append(text_para)
        break

print(final_outputs)        

Output

['Cred că fac multe greșeli.']