How Instruction and Reasoning Data shape Post-Training: Data Quality through the Lens of Layer-wise Gradients
By: Ming Li , Yanhong Li , Ziyue Li and more
Potential Business Impact:
Helps AI learn better from good or bad examples.
As the post-training of large language models (LLMs) advances from instruction-following to complex reasoning tasks, understanding how different data affect finetuning dynamics remains largely unexplored. In this paper, we present a spectral analysis of layer-wise gradients induced by low/high-quality instruction and reasoning data for LLM post-training. Our analysis reveals that widely-studied metrics for data evaluation, e.g., IFD, InsTag, Difficulty, and Reward, can be explained and unified by spectral properties computed from gradients' singular value decomposition (SVD). Specifically, higher-quality data are usually associated with lower nuclear norms and higher effective ranks. Notably, effective rank exhibits better robustness and resolution than nuclear norm in capturing subtle quality differences. For example, reasoning data achieves substantially higher effective ranks than instruction data, implying richer gradient structures on more complex tasks. Our experiments also highlight that models within the same family share similar gradient patterns regardless of their sizes, whereas different model families diverge significantly. Providing a unified view on the effects of data quality across instruction and reasoning data, this work illuminates the interplay between data quality and training stability, shedding novel insights into developing better data exploration strategies for post-training.
Similar Papers
Understanding Post-Training Structural Changes in Large Language Models
Machine Learning (CS)
Changes how AI learns, making it more predictable.
Influence Functions for Efficient Data Selection in Reasoning
Machine Learning (CS)
Teaches computers to think better with fewer examples.
Q-Ponder: A Unified Training Pipeline for Reasoning-based Visual Quality Assessment
CV and Pattern Recognition
Helps computers judge picture quality better.