Long-context language understanding on LV-Eval
7.05CMRC (Mixup)Qwen3-1.7B-Mamba
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-1.7B-MambaAvg. Len.=128k, Mechanism=Mamba-enhanced memory model2026.02 | 7.05 | 10.64 | 1 | 0.5 | 2.73 | 5.11 | 2.52 | 1.66 | 4.25 | 5.81 | 6.72 | 4.36 | |
| Qwen3-1.7BAvg. Len.=128k, Mechanism=Full attention2026.02 | 6.11 | 11.97 | 1.8 | 0.38 | 4.74 | 6.01 | 3.26 | 1.72 | 6.6 | 7.74 | 7.86 | 5.29 | |
| Qwen3-1.7B-ALLMEMAvg. Len.=128k, Mechanism=ALLMEM memory model2026.02 | 5.64 | 9.91 | 6.83 | 7 | 6.81 | 4.61 | 2.24 | 1.14 | 4.42 | 5.78 | 6.8 | 5.56 | |
| Qwen3-1.7B-SWA-SinksAvg. Len.=128k, Mechanism=Sliding-window attention, Sliding window length=8k, Attention sinks=1282026.02 | 5.27 | 9.46 | 1 | 0.31 | 4.96 | 2.83 | 1.71 | 1.34 | 4.61 | 4.77 | 5.14 | 3.76 |