Core Concepts
Performance of dense retrieval models follows precise power-law scaling related to model size and data size.
Abstract
This study explores scaling laws in dense retrieval models, focusing on model size and data size. The core findings include the power-law scaling relationship observed in model performance. Different data augmentation methods and potential applications of the scaling law are also discussed.
Introduction
Scaling laws in language data.
Neural scaling laws.
Dense retrieval models.
Model Size Scaling
Performance improves with larger model sizes.
Contrastive perplexity follows a power-law scaling.
Fitting parameters for model size scaling.
Data Size Scaling
Performance scales with data size.
Contrastive perplexity follows a power-law scaling.
Fitting parameters for data size scaling.
Annotation Quality
Different annotation qualities impact performance.
Weak supervision vs. human annotations.
Potential of LLM-based data augmentation.
Application in Budget Allocation
Predicted contrastive perplexity under different cost budgets.
Optimal model size based on budget constraints.
Consideration of inference costs.
Conclusions and Future Work
Power-law scaling in dense retrieval models.
Limitations and future research directions.
Stats
Results indicate that the contrastive perplexity follows a power-law scaling in relation to model size and data size.
Quotes
"The performance of dense retrieval models follows a precise power-law scaling related to the model size and the number of annotations."