BART-SLED (SLiding-Encoder and Decoder, large-sized model)

SLED models use pretrained, short-range encoder-decoder models, and apply them over long-text inputs by splitting the input into multiple overlapping chunks, encoding each independently and perform fusion-in-decoder

Model description

This SLED model is based on the BART model, which is described in its model card. BART is particularly effective when fine-tuned for text generation (e.g. summarization, translation) but also works well for comprehension tasks (e.g. text classification, question answering). When used as a BART-SLED model, it can be applied on long text tasks.

This model was finetuned on the GovReport

Intended uses & limitations

You can use the raw model for text infilling. However, the model is mostly meant to be fine-tuned on a supervised dataset.

How to use

To use the model, you first need to install py-sled in your environment (or clone the code from the official repository)

pip install py-sled

For more installation instructions, see here.

Once installed, SLED is fully compatible with HuggingFace's AutoClasses (AutoTokenizer, AutoConfig, AutoModel and AutoModelForCausalLM) and can be loaded using the from_pretrained methods

import sled   # *** required so that SledModels will be registered for the AutoClasses ***
model = AutoModel.from_pretrained('tau/bart-large-sled')

Here is how to use this model in PyTorch:

from sled import SledTokenizer, SledModel
tokenizer = SledTokenizer.from_pretrained('tau/bart-large-sled')
model = SledModel.from_pretrained('tau/bart-large-sled')
inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
outputs = model(**inputs)
last_hidden_states = outputs.last_hidden_state

You can also replace SledModel by SledModelForConditionalGeneration for Seq2Seq generation

model = SledModelForConditionalGeneration.from_pretrained('tau/bart-large-sled')

In case you wish to apply SLED on a task containing a prefix (e.g. question) which should be given as a context to every chunk, you can pass the prefix_length tensor input as well (A LongTensor in the length of the batch size).

import torch
import sled   # *** required so that SledModels will be registered for the AutoClasses ***
tokenizer = AutoTokenizer.from_pretrained('tau/bart-large-sled')
model = AutoModel.from_pretrained('tau/bart-large-sled')
document_input_ids = tokenizer("Dogs are great for you.", return_tensors="pt").input_ids
prefix_input_ids = tokenizer("Are dogs good for you?", return_tensors="pt").input_ids
input_ids = torch.cat((prefix_input_ids, document_input_ids), dim=-1)
attention_mask = torch.ones_like(input_ids)
prefix_length = torch.LongTensor([[prefix_input_ids.size(1)]])

outputs = model(input_ids=input_ids, attention_mask=attention_mask, prefix_length=prefix_length)
last_hidden_states = outputs.last_hidden_state

BibTeX entry and citation info

Please cite both the SLED paper and the BART paper by Lewis et al as well as GovReport by Huang et al

@inproceedings{Ivgi2022EfficientLU,
  title={Efficient Long-Text Understanding with Short-Text Models},
  author={Maor Ivgi and Uri Shaham and Jonathan Berant},
  year={2022}
}

@article{DBLP:journals/corr/abs-1910-13461,
  author    = {Mike Lewis and
               Yinhan Liu and
               Naman Goyal and
               Marjan Ghazvininejad and
               Abdelrahman Mohamed and
               Omer Levy and
               Veselin Stoyanov and
               Luke Zettlemoyer},
  title     = {{BART:} Denoising Sequence-to-Sequence Pre-training for Natural Language
               Generation, Translation, and Comprehension},
  journal   = {CoRR},
  volume    = {abs/1910.13461},
  year      = {2019},
  url       = {http://arxiv.org/abs/1910.13461},
  eprinttype = {arXiv},
  eprint    = {1910.13461},
  timestamp = {Thu, 31 Oct 2019 14:02:26 +0100},
  biburl    = {https://dblp.org/rec/journals/corr/abs-1910-13461.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

@inproceedings{huang2021govreport,
    title = "Efficient Attentions for Long Document Summarization",
    author = "Huang, Luyang  and
      Cao, Shuyang  and
      Parulian, Nikolaus  and
      Ji, Heng  and
      Wang, Lu",
    booktitle = "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jun,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.naacl-main.112",
    doi = "10.18653/v1/2021.naacl-main.112",
    pages = "1419--1436"
}