Suffix completion under prefix compression on FineWeb held-out
0.259∆PPLQWEN2.5-1.5B KV-CAT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| QWEN2.5-1.5B KV-CATModel=QWEN2.5-1.5B, Keep Ratio=40%, Variant=KV-CAT2026.05 | 0.259 | 0.0326 | 92.4 | |
| QWEN2.5-1.5B KV-CATModel=QWEN2.5-1.5B, Keep Ratio=5%, Variant=KV-CAT2026.05 | 0.292 | 0.0361 | 90.9 | |
| QWEN2.5-1.5B KV-CATModel=QWEN2.5-1.5B, Keep Ratio=10%, Variant=KV-CAT2026.05 | 0.32 | 0.0366 | 90.7 | |
| QWEN2.5-0.5B KV-CATModel=QWEN2.5-0.5B, Keep Ratio=40%, Variant=KV-CAT2026.05 | 0.36 | 0.0272 | 92.5 | |
| QWEN2.5-0.5B KV-CATModel=QWEN2.5-0.5B, Keep Ratio=10%, Variant=KV-CAT2026.05 | 0.422 | 0.0321 | 91.4 | |
| QWEN2.5-0.5B KV-CATModel=QWEN2.5-0.5B, Keep Ratio=20%, Variant=KV-CAT2026.05 | 0.438 | 0.0374 | 90.8 | |
| QWEN2.5-0.5B KV-CATModel=QWEN2.5-0.5B, Keep Ratio=5%, Variant=KV-CAT2026.05 | 0.461 | 0.0385 | 91.1 | |
| QWEN2.5-1.5B KV-CATModel=QWEN2.5-1.5B, Keep Ratio=20%, Variant=KV-CAT2026.05 | 0.536 | 0.0491 | 90.2 | |
| QWEN2.5-1.5B BaseModel=QWEN2.5-1.5B, Keep Ratio=5%, Variant=Base2026.05 | 0.573 | 0.057 | 89.2 | |
| QWEN2.5-0.5B BaseModel=QWEN2.5-0.5B, Keep Ratio=40%, Variant=Base2026.05 | 0.592 | 0.0431 | 90.3 | |
| QWEN2.5-0.5B BaseModel=QWEN2.5-0.5B, Keep Ratio=10%, Variant=Base2026.05 | 0.662 | 0.0483 | 89.3 | |
| QWEN2.5-1.5B BaseModel=QWEN2.5-1.5B, Keep Ratio=40%, Variant=Base2026.05 | 0.742 | 0.0653 | 90 | |
| QWEN2.5-0.5B BaseModel=QWEN2.5-0.5B, Keep Ratio=20%, Variant=Base2026.05 | 0.777 | 0.0549 | 88.7 | |
| QWEN2.5-0.5B BaseModel=QWEN2.5-0.5B, Keep Ratio=5%, Variant=Base2026.05 | 0.78 | 0.0562 | 88.5 | |
| QWEN2.5-1.5B BaseModel=QWEN2.5-1.5B, Keep Ratio=10%, Variant=Base2026.05 | 1.027 | 0.0796 | 88.3 | |
| QWEN2.5-1.5B BaseModel=QWEN2.5-1.5B, Keep Ratio=20%, Variant=Base2026.05 | 1.093 | 0.0876 | 87.7 |