The viewer is disabled because this dataset repo requires arbitrary Python code execution. Please consider removing the loading script and relying on automated data support (you can use convert_to_parquet from the datasets library). If this is not possible, please open a discussion for direct help.

MediaSum dataset for summarization

Summarization dataset copied from MediaSum: A Large-scale Media Interview Dataset for Dialogue Summarization

This dataset is compatible with the run_summarization.py script from Transformers if you add this line to the summarization_name_mapping variable:

"ccdv/mediasum": ("document", "summary")

Configs

4 possibles configs:

  • roberta will concatenate documents with "</s>"
  • newline will concatenate documents with "\n"
  • bert will concatenate documents with "[SEP]"
  • list will return the list of documents instead of a single string

Add _prepended to config name to prepend the speaker name before each dialogue: speaker: text
Default is roberta_prepended (compatible with BART).

Data Fields

  • id: paper id
  • document: a string/list containing the body of a set of documents
  • summary: a string containing the abstract of the set

Data Splits

This dataset has 3 splits: train, validation, and test. \

Dataset Split Number of Instances
Train 443596
Validation 10000
Test 10000

Cite original article

@article{zhu2021mediasum,
  title={MediaSum: A Large-scale Media Interview Dataset for Dialogue Summarization},
  author={Zhu, Chenguang and Liu, Yang and Mei, Jie and Zeng, Michael},
  journal={arXiv preprint arXiv:2103.06410},
  year={2021}
}
Downloads last month
171

Models trained or fine-tuned on ccdv/mediasum