WISCA: A Lightweight Model Transition Method to Improve LLM Training via Weight Scaling

TL;DR

WISCA is a lightweight weight scaling method that improves training efficiency and model quality of Transformer-based LLMs by optimizing weight patterns without altering network architecture.

Contribution

The paper introduces WISCA, a novel weight scaling technique that enhances LLM training by systematically improving weight patterns during training.

Findings

WISCA achieves a 5.6% average improvement on zero-shot validation tasks.

WISCA reduces training perplexity by 2.12% on average.

WISCA significantly improves convergence quality in GQA architectures and LoRA fine-tuning.

Abstract

Transformer architecture gradually dominates the LLM field. Recent advances in training optimization for Transformer-based large language models (LLMs) primarily focus on architectural modifications or optimizer adjustments. However, these approaches lack systematic optimization of weight patterns during training. Weight pattern refers to the distribution and relative magnitudes of weight parameters in a neural network. To address this issue, we propose a Weight Scaling method called WISCA to enhance training efficiency and model quality by strategically improving neural network weight patterns without changing network structures. By rescaling weights while preserving model outputs, WISCA indirectly optimizes the model's training trajectory. Experiments demonstrate that WISCA significantly improves convergence quality (measured by generalization capability and loss reduction), particularly in LLMs with Grouped Query Attention (GQA) architectures and LoRA fine-tuning tasks. Empirical results show 5.6% average improvement on zero-shot validation tasks and 2.12% average reduction in training perplexity across multiple architectures.