Token Compression on PangolinBench
92.2Compression Ratio (Formal TC)VoxCPM2
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| VoxCPM2Vocabulary size=122,7532026.07 | 92.2 | 97.9 | 0.982 | 70.3 | 36.1 | 1.364 | 45.7 | 46.8 | 61 | |
| LLaMA 3.1Vocabulary size=128,2562026.07 | 90.2 | 91.7 | 1 | 63.7 | 33.5 | 1.205 | 42.4 | 40.5 | 55.4 | |
| GPT-4oVocabulary size=200,0192026.07 | 90.2 | 97.9 | 1.009 | 61.5 | 31.9 | 1.227 | 45.7 | 40.5 | 55.8 | |
| Qwen 2.5/3Vocabulary size=151,6692026.07 | 73.5 | 79.2 | 0.919 | 69.2 | 34.6 | 1.114 | 45.7 | 44.7 | 55.9 | |
| Gemma 3/4Vocabulary size=262,1442026.07 | 71.6 | 75 | 0.892 | 58.2 | 31.4 | 0.795 | 38 | 42.8 | 52.6 | |
| PangolinVocabulary size=114,8222026.07 | 67.6 | 70.8 | 0.73 | 71.4 | 37.2 | 1.068 | 47.8 | 40 | 48.5 | |
| Qwen 3.6Vocabulary size=248,0702026.07 | 67.6 | 75 | 0.757 | 61.5 | 34 | 1.386 | 41.3 | 44.1 | 54.1 | |
| TAIDEVocabulary size=318,0802026.07 | 55.9 | 68.8 | 0.622 | 53.8 | 29.3 | 0.75 | 38 | 42.3 | 48.6 |