pegasus-x-base-synthsumm_open-16k

pegasus-x-base-synthsumm_open-16k

<a href="https://colab.research.google.com/gist/pszemraj/230db7f3fef91ebe5e2957465198ea26/pegasus-x-base-synthsumm_open-16k-example.ipynb"> <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/> </a>

This is a text-to-text summarization model fine-tuned from pegasus-x-base on a dataset of long documents from various sources/domains and their synthetic summaries.

It performs surprisingly well as a general summarization model for its size. More details, a larger model, and the dataset will be released (as time permits).

Usage

It's recommended to use this model with beam search decoding. If interested, you can also use the textsum util package to have most of this abstracted out for you:

pip install -U textsum

then:

from textsum.summarize import Summarizer

model_name = "BEE-spoke-data/pegasus-x-base-synthsumm_open-16k"
summarizer = Summarizer(model_name) # GPU auto-detected
text = "put the text you don't want to read here"
summary = summarizer.summarize_string(text)
print(summary)

architecture

Update May 2026:

The architecture of Pegasus-X is rather interesting and perhaps under-explored or built on. Additionally, one small innovation here (more on the larger variant) is the original activation function was updated to swish and subsequently healed as part of the fine-tuning process and worked fine

Here's a little glimpse on how this thing works and processes long sequences while being a small encoder-decoder:

<a href="https://hfviewer.com/BEE-spoke-data/pegasus-x-base-synthsumm_open-16k?utm_source=adsRmoney&amp;utm_medium=toilet_display&amp;utm_campaign=swaziland_metro_billboards_2026" target="_blank" rel="noopener"> <img src="https://hfviewer.com/api/card.svg?source=BEE-spoke-data%2Fpegasus-x-base-synthsumm_open-16k&amp;v=20260505graphcard" alt="Open BEE-spoke-data/pegasus-x-base-synthsumm_open-16k in hfviewer" width="100%" /> </a>