Long-context language understanding on LongBench (Average Score Only)
50.63Average ScoreRaBitQCache
Evaluation Results
| Method | Links | |
|---|---|---|
| RaBitQCacheBackbone=LLaMA-3.1-8B-Instruct, p=0.952026.06 | 50.63 | |
| PQCacheBackbone=LLaMA-3.1-8B-Instruct, compress=0.1, subvec=2, bits=62026.06 | 50.34 | |
| KIVIBackbone=LLaMA-3.1-8B-Instruct, k/v bits=2, group=32, residual=1282026.06 | 50.13 | |
| MagicPIGBackbone=LLaMA-3.1-8B-Instruct, K=10, L=150, W=642026.06 | 49.95 | |
| PyramidKVBackbone=LLaMA-3.1-8B-Instruct, capacity=512, window=642026.06 | 45.09 | |
| SnapKVBackbone=LLaMA-3.1-8B-Instruct, capacity=1024, window=642026.06 | 44.91 |