List your company

Tokenizer Evaluation on mC4 unbalanced 30-lang

0.0296Compression RatePA-BPE (window+hybrid-127k)
0.0188880.0216690.024450.027231Aug 6, 2025

Evaluation results

MethodLinks
PA-BPE (window+hybrid-127k)Vocabulary Size=127k2025.080.02960.123149.33.6980.460.0320.530.596
BPEVocabulary Size=256k2025.080.02940.1248.33.6960.460.0560.5290.598
PA-BPE (window+hybrid)Vocabulary Size=256k2025.080.02940.1217493.7090.460.030.5240.59
PA-BPE (hybrid-127k)Vocabulary Size=127k2025.080.02940.118847.93.7220.460.0270.5210.59
PA-BPE (window)Vocabulary Size=256k2025.080.02920.120348.93.7250.470.0320.5140.58
PA-BPE (hybrid)Vocabulary Size=256k2025.080.02910.115747.13.7390.470.0260.5130.582
PA-BPEVocabulary Size=256k2025.080.02880.1143473.7520.470.0280.5010.57
PA-BPE (window+hybrid)Vocabulary Size=128k2025.080.02660.076167.94.1270.490.0310.430.517
PA-BPE (hybrid)Vocabulary Size=128k2025.080.02640.073766.34.1160.50.0270.4260.515
BPEVocabulary Size=128k2025.080.02620.071865.14.1780.50.0670.4340.527
PA-BPE (window)Vocabulary Size=128k2025.080.02620.074967.94.1520.50.0290.4170.507
PA-BPEVocabulary Size=128k2025.080.0260.072766.44.1380.50.0270.4180.508
UnigramLMVocabulary Size=128k2025.080.01930.045756.34.5360.290.0950.1470.262