AIR-hl/

helpsteer3_preference

This is a binarized preference datasets from nvidia/HelpSteer3. HelpSteer3 contains 40,476 preference samples, each containing a domain, language, context, two responses, an overall preference score between the responses as well as individual preferences from up to 3 annotators. Each individual preference contains a preference score in addition to a concise reasoning for their preference in 1-2 sentences. Data is split into 95% train and 5% validation. I processed the… See the full description on the dataset page: https://huggingface.co/datasets/AIR-hl/helpsteer3_preference.