Passkey Retrieval on Passkey Retrieval
100Retrieval Success Rate (4K)fp16
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| fp16Backbone=LLaMA-2-7B-32K2024.01 | 100 | 100 | 100 | 100 | 100 | 16 | |
| nuq4-1%Backbone=LLaMA-2-7B-32K, Quantization Type=Non-Uniform Quantization, Outlier Ratio=1%2024.01 | 100 | 100 | 100 | 100 | 100 | 4.33 | |
| nuq3-1%Backbone=LLaMA-2-7B-32K, Quantization Type=Non-Uniform Quantization, Outlier Ratio=1%2024.01 | 100 | 98 | 100 | 100 | 100 | 3.33 | |
| nuq2-1%Backbone=LLaMA-2-7B-32K, Quantization Type=Non-Uniform Quantization, Outlier Ratio=1%2024.01 | 100 | 100 | 98 | 100 | 100 | 2.33 | |
| fp16Backbone=Llama-2-70B-32K2024.01 | 100 | 100 | 100 | 100 | 100 | 16 | |
| nuq4-1%Backbone=Llama-2-70B-32K, Quantization Type=Non-Uniform Quantization, Outlier Ratio=1%2024.01 | 100 | 100 | 100 | 100 | 100 | 4.35 | |
| nuq3-1%Backbone=Llama-2-70B-32K, Quantization Type=Non-Uniform Quantization, Outlier Ratio=1%2024.01 | 100 | 100 | 100 | 100 | 100 | 3.35 | |
| nuq2-1%Backbone=Llama-2-70B-32K, Quantization Type=Non-Uniform Quantization, Outlier Ratio=1%2024.01 | 98 | 98 | 96 | 100 | 74 | 2.35 | |
| StateXArchitecture=GLA2025.09 | 93 | — | 77 | 34 | 6 | — | |
| VanillaArchitecture=GLA2025.09 | 74 | — | 41 | 13 | 1 | — | |
| KIVI-2-gs32-r128Backbone=LLaMA-2-7B-32K, Bits=2, Group Size=32, Residual=128-element fp162024.01 | 72 | 76 | 72 | 68 | 70 | 3.05 |