Roberta2Roberta_L-24_gigaword EncoderDecoder model

The model was introduced in this paper by Sascha Rothe, Shashi Narayan, Aliaksei Severyn and first released in this repository.

The model is an encoder-decoder model that was initialized on the roberta-large checkpoints for both the encoder and decoder and fine-tuned on headline generation using the Gigaword dataset, which is linked above.

Disclaimer: The model card has been written by the Hugging Face team.

How to use

You can use this model for extreme summarization, e.g.

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("google/roberta2roberta_L-24_gigaword")
model = AutoModelForSeq2SeqLM.from_pretrained("google/roberta2roberta_L-24_gigaword")

article = """australian shares closed down #.# percent monday
following a weak lead from the united states and
lower commodity prices , dealers said ."""

input_ids = tokenizer(article, return_tensors="pt").input_ids
output_ids = model.generate(input_ids)[0]
print(tokenizer.decode(output_ids, skip_special_tokens=True))
# should output
# australian shares close down #.# percent.