Open nmattei opened 1 year ago
What's your idea here? Like how do you convert it into PrefLib?
Similar idea: https://huggingface.co/datasets/Anthropic/hh-rlhf and https://huggingface.co/datasets/OpenAssistant/oasst1
For Anthropic it does not seem to be good since each answer is presented only once, so we would just have highly incomplete data.
For Oasst I'm still trying to understand the file...
Same for OASST actually, each message seems unique.
Q/A Response dataset? https://huggingface.co/datasets/stanfordnlp/SHP