Tending Towards Stability: Convergence Challenges in Small Language Models
This addresses inefficiencies in small language models for researchers and practitioners seeking cost-effective AI solutions, but it is incremental as it analyzes known issues without proposing a new solution.
The paper investigates why smaller language models underperform larger ones during late pretraining, finding that layers in smaller models converge slower and less stably, especially when parameters have lower effective rank, compared to larger models that stabilize within the first 20% of training.
Increasing the number of parameters in language models is a common strategy to enhance their performance. However, smaller language models remain valuable due to their lower operational costs. Despite their advantages, smaller models frequently underperform compared to their larger counterparts, even when provided with equivalent data and computational resources. Specifically, their performance tends to degrade in the late pretraining phase. This is anecdotally attributed to their reduced representational capacity. Yet, the exact causes of this performance degradation remain unclear. We use the Pythia model suite to analyse the training dynamics that underlie this phenomenon. Across different model sizes, we investigate the convergence of the Attention and MLP activations to their final state and examine how the effective rank of their parameters influences this process. We find that nearly all layers in larger models stabilise early in training - within the first 20% - whereas layers in smaller models exhibit slower and less stable convergence, especially when their parameters have lower effective rank. By linking the convergence of layers' activations to their parameters' effective rank, our analyses can guide future work to address inefficiencies in the learning dynamics of small models.