GXSum is an English news summarization dataset that provides LLM-generated, high-quality reference summaries aligned to the XSum (Extreme Summarization) benchmark. GXSum is designed to study how improving reference quality affects training and evaluation of classic summarization models, and to enable apples-to-apples comparisons between classic models and modern LLMs.
This dataset contains three variants (available as configs/subsets in the same… See the full description on the dataset page:
https://huggingface.co/datasets/GlycerinLOL/GXSum.
