R-Select
A Robust Multi-Metric Data Selection Approach for Fine-Tuning Large Language Models
Dataset Performance vs. dataset size.
R-Select-100k is a compact post-training dataset selected by R-Select, a robust multi-metric data selection framework accepted by KDD 2026. It is curated from a heterogeneous instruction-tuning pool of over 3.4M samples and selects 100K high-value samples by optimizing 30 data-quality metrics through hierarchical Bayesian optimization.… See the full description on the dataset page:
https://huggingface.co/datasets/GX-XinGao/R-Select-100k.