As vocabulary size grows, the optimal embedding-to-hidden learning rate ratio in LLMs scales as Θ(√width), a regime the authors call the Large Vocabulary parameterization.
Alemi, Roman Novak, Peter J
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Optimal Embedding Learning Rate in LLMs: The Effect of Vocabulary Size
As vocabulary size grows, the optimal embedding-to-hidden learning rate ratio in LLMs scales as Θ(√width), a regime the authors call the Large Vocabulary parameterization.