Long-context Understanding on RULER 64K context 1.0 (test)
100N-S Score 1Dense
Evaluation Results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DenseModel=Qwen2.5-7B-Instruct-1M, Context Length=64K, K.V. BW (× ↓)=1.002025.03 | 100 | 100 | 100 | 100 | 100 | 95.83 | 84.38 | 60.42 | 90.63 | 86.81 | 91.81 | |
| Single SVDModel=Qwen2.5-7B-Instruct-1M, Context Length=64K, K.V. BW (× ↓)=8.402025.03 | 100 | 97.92 | 96.88 | 98.96 | 97.4 | 91.15 | 64.58 | 56.25 | 73.75 | 61.46 | 83.84 | |
| xKVModel=Qwen2.5-7B-Instruct-1M, Context Length=64K, K.V. BW (× ↓)=8.032025.03 | 100 | 100 | 100 | 98.96 | 100 | 90.63 | 80.21 | 58.33 | 82.08 | 81.94 | 89.22 | |
| SnapKVModel=Qwen2.5-7B-Instruct-1M, Context Length=64K, K.V. BW (× ↓)=8.002025.03 | 100 | 96.88 | 97.92 | 31.25 | 95.31 | 83.07 | 84.38 | 59.38 | 91.25 | 80.56 | 82 | |
| PyramidKVModel=Qwen2.5-7B-Instruct-1M, Context Length=64K, K.V. BW (× ↓)=8.002025.03 | 100 | 93.75 | 96.88 | 16.67 | 90.37 | 80.73 | 84.38 | 59.38 | 89.17 | 76.39 | 78.77 | |
| TokenButlerModel=Qwen2.5-7B-Instruct-1M, Context Length=64K, K.V. BW (× ↓)=8.002025.03 | 100 | 100 | 100 | 96.88 | 99.47 | 90.63 | 84.38 | 58.33 | 89.17 | 82.64 | 90.15 | |
| MiniCacheModel=Qwen2.5-7B-Instruct-1M, Context Length=64K, K.V. BW (× ↓)=1.302025.03 | 26.04 | 0 | 0 | 0 | 0 | 0 | 12.5 | 14.58 | 0.42 | 3.47 | 5.7 | |
| StreamingLLMModel=Qwen2.5-7B-Instruct-1M, Context Length=64K, K.V. BW (× ↓)=8.002025.03 | 15.63 | 12.5 | 12.5 | 9.38 | 14.84 | 17.71 | 46.88 | 43.75 | 13.13 | 89.24 | 27.56 | |
| KIVIModel=Qwen2.5-7B-Instruct-1M, Context Length=64K, K.V. BW (× ↓)=7.102025.03 | 0 | 2.08 | 3.13 | 13.54 | 0 | 0.78 | 48.96 | 43.75 | 36.46 | 40.63 | 18.93 |