---
license: apache-2.0
task_categories:
- text2text-generation
language:
- en
---
The data is orginally from Standford SHP dataset. I build two dataset: hard and soft based on the BERT-Large embedding of CLS. 
Hard is the most similiar pairs and soft is the most dis-similiar pairs. To make the two dataset different, I also filter out the overlaped pairs (since some prompt only has one pair).
The prompt distribution is not uniform, so I only selected one pair for each prompt for hard and soft dataset.