You can deploy this model and start earning money today!

josh-oo/
custom-decoder-ats

Name: custom-decoder-ats
Rating: 2.5 (13 reviews)

Views

No views yet

German text simplification with custom decoder

This model was initialized from an mBART model and the decoder was replaced by a GPT2 language model pre-trained for German Easy Language. For more details, visit our Github repository.

Usage

python

import torch
from transformers import AutoTokenizer
from transformers import AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("josh-oo/custom-decoder-ats")

##gerpt

#model = AutoModelForSeq2SeqLM.from_pretrained("josh-oo/custom-decoder-ats", trust_remote_code=True, revision="35197269f0235992fcc6b8363ca4f48558b624ff")
#decoder_tokenizer = AutoTokenizer.from_pretrained("josh-oo/gerpt2")

##dbmdz

model = AutoModelForSeq2SeqLM.from_pretrained("josh-oo/custom-decoder-ats", trust_remote_code=True, revision="4accedbe0b57d342d95ff546b6bbd3321451d504")
decoder_tokenizer = AutoTokenizer.from_pretrained("josh-oo/german-gpt2-easy")
decoder_tokenizer.add_tokens(['<</s>>','<<s>>','<<pad>>'])

##

example_text = "In tausenden Schweizer Privathaushalten kümmern sich Haushaltsangestellte um die Wäsche, betreuen die Kinder und sorgen für Sauberkeit. Durchschnittlich bekommen sie für die Arbeit rund 30 Franken pro Stunde Bruttolohn. Der grösste Teil von ihnen erhält aber 28 Franken."

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)
model.eval()

test_input = tokenizer([example_text], return_tensors="pt", padding=True, pad_to_multiple_of=1024)
for key, value in test_input.items():
  test_input[key] = value.to(device)

outputs = model.generate(**test_input, num_beams=3, max_length=1024)
decoder_tokenizer.batch_decode(outputs)

Citation

If you use our mode, please cite:
@misc{anschütz2023language,
title={Language Models for German Text Simplification: Overcoming Parallel Data Scarcity through Style-specific Pre-training},
author={Miriam Anschütz and Joshua Oehms and Thomas Wimmer and Bartłomiej Jezierski and Georg Groh},
year={2023},
eprint={2305.12908},
archivePrefix={arXiv},
primaryClass={cs.CL}
}