Distill Pegasus Xsum 16 4

🚀 Pegasus Models

Pegasus is a model designed for the task of summarization. This README provides details about the model, its training, and experimental results.

🚀 Quick Start

For detailed documentation of Pegasus models, please refer to here. The original TF 1 code can be found here.

✨ Features

• Task: Summarization
• Authors: Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu on Dec 18, 2019
• Maintainer: @sshleifer

📚 Documentation

Mixed & Stochastic Checkpoints

We trained a Pegasus model with sampled gap sentence ratios on both C4 and HugeNews, and stochastically sampled important sentences. The updated results are reported in the following table:

The "Mixed & Stochastic" model has the following changes:

• Trained on both C4 and HugeNews (dataset mixture is weighted by their number of examples).
• Trained for 1.5M instead of 500k (we observe slower convergence on pretraining perplexity).
• The model uniformly samples a gap sentence ratio between 15% and 45%.
• Importance sentences are sampled using a 20% uniform noise to importance scores.
• The SentencePiece tokenizer is updated to be able to encode newline characters.

(*) The numbers of Wikihow and BigPatent datasets are not comparable because of changes in tokenization and data:

• The Wikihow dataset contains newline characters which are useful for paragraph segmentation. The C4 and HugeNews model's SentencePiece tokenizer doesn't encode newline and loses this information.
• We updated the BigPatent dataset to preserve casing, and some format cleanings are also changed. Please refer to changes in TFDS.

The "Mixed & Stochastic" model has the following changes (from pegasus - large in the paper):

• Trained on both C4 and HugeNews (dataset mixture is weighted by their number of examples).
• Trained for 1.5M instead of 500k (we observe slower convergence on pretraining perplexity).
• The model uniformly samples a gap sentence ratio between 15% and 45%.
• Importance sentences are sampled using a 20% uniform noise to importance scores.
• The SentencePiece tokenizer is updated to be able to encode newline characters.

Citation

@misc{zhang2019pegasus,
    title={PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization},
    author={Jingqing Zhang and Yao Zhao and Mohammad Saleh and Peter J. Liu},
    year={2019},
    eprint={1912.08777},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}