Addressing Training Instability in a Large Language Model

Match each architectural modification with its primary purpose in stabilizing the training of very large language models.

A team training a multi-billion parameter language model observes that the training process frequently fails due to sudden, large spikes in the loss function and exploding gradient values. This instability becomes more pronounced as they increase the model's depth. Which of the following architectural modifications is most specifically designed to counteract this particular problem and improve training stability in very deep models?

Key Issues in Large-Scale LLM Training

A research lab is pre-training a new language model with billions of parameters on a petabyte-scale dataset. Midway through the process, they observe that the model's learning progress becomes highly erratic, and the training process frequently crashes. Which statement best analyzes the fundamental challenge they are facing?

Distributed Training for Large-Scale LLMs

Scaling Laws for LLMs

During the pre-training phase of a large language model, consistently increasing the volume of the training data and the number of model parameters will reliably lead to a more stable training process and better performance.

LLM Pre-training Strategy Analysis

Data Demand for Large Language Models