Beyond Text Compression: Evaluating Tokenizers Across Scales
令牌设计师的设计显着影响语言模型性能，但是评估令牌质量仍然具有挑战性。尽管文本压缩已成为一种常见的内在度量，但最近的工作质疑其作为质量指标的可靠性。 We investigate whether evaluating tokenizers on smaller models (350M parameters) reliably predicts their impact at larger scales (2.7B parameters).Through experiments with established tokenizers from widely-adopted language m
