This repository provides a scored SFT dataset collection currently featuring 63 high-quality instruction-following datasets with nearly 25 million samples. The core value lies in its 30-dimensional scoring: every sample has been evaluated on metrics such as IFD, PPL, Deita_Quality, and 27 others, enabling fine-grained data selection for filtering, curriculum learning, and mixture optimization.
Key features:
30 metrics per sample — From lexical… See the full description on the dataset page:
https://huggingface.co/datasets/OpenDataArena/OpenDataArena-scored-data-2603.
