ConceptChinchilla Scaling Law

Chinchilla Law

A training efficiency theory proposed by DeepMind in 2022 that revises the optimal compute allocation ratios from earlier scaling law research. It states that dataset size and model parameter count should be scaled in equal proportion for optimal training efficiency.

Core Facts

The Chinchilla law states that under a fixed compute budget, appropriately reducing model size and increasing training data volume often yields better performance.

Verified80% confidence4 sources

Source Articles

Rethinking LLM Scaling Laws: How to Get the Token-to-Parameter Ratio RightOct 7 AI's Second Half: The Decisive Battle Shifts from Models to InfrastructureJul 10