Description
How to select, mix and weight data for LLM training varies across papers, making fair comparison hard. DataFlex is a unified benchmark and evaluation platform for data-centric LLM training that compares data selection, mixing and reweighting in one framework.
It standardizes experiments and evaluation, published at NeurIPS 2026.
Selection:Methods compared.
Mixing:Strategies evaluated.
Reweighting:Sample weights.
Unified:Fair comparison.
It standardizes experiments and evaluation, published at NeurIPS 2026.
Features
Selection:Methods compared.
Mixing:Strategies evaluated.
Reweighting:Sample weights.
Unified:Fair comparison.
