This is a processed subset of the openai/summarize_from_feedback comparisons subset, including training and validation splits.
The original dataset consists of paired human comparisons between summary candidates for given source texts. This processed version aggregates all comparisons per unique text to determine the overall best (chosen) and worst (rejected) summaries using the Bradley-Terry model.
