Hierarchical Data Selection via Manifold Coverage and Sparse Feature Coverage in LLM Post-training
Hierarchical data selection methodology optimizes LLM training by balancing semantic grouping and sparse feature coverage.
As supervised fine-tuning data continues to scale, selecting high-value subsets from large candidate pools is crucial for reducing training cost and improving model performance. Existing methods often measure diversity directly in the original embedding space, where geometric met...
ArXiv CS.LG
Source
Generated insight by OnAirToday AI
This research introduces a hierarchical approach to data selection that disentangles global semantic structures from fine-grained features, significantly improving training efficiency. By refining how models consume data, this method achieves full-training performance using only a fraction of the original dataset.
Comments
Keep comments relevant and respectful. No links, advertising, self-promotion, or spam.