Long-context language understanding on LongBench Zh
62.24Single Document PerformanceSentinel
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| SentinelProxy Model=Qwen2.5-0.5B, Context Constraint=2K, Downstream LLM=Qwen2.5-Instruct-7B2025.05 | 62.24 | 18.57 | 40.41 | 38.02 | |
| Original PromptDownstream LLM=Qwen2.5-Instruct-7B2025.05 | 60.06 | 18.21 | 39.14 | 37.3 | |
| Raw AttentionProxy Model=Qwen2.5-0.5B, Context Constraint=2K, Downstream LLM=Qwen2.5-Instruct-7B2025.05 | 51.72 | 17.29 | 34.5 | 33.12 | |
| RandomContext Constraint=2K, Downstream LLM=Qwen2.5-Instruct-7B2025.05 | 43.18 | 17.22 | 30.2 | 28.3 | |
| EmptyContext Constraint=2K, Downstream LLM=Qwen2.5-Instruct-7B2025.05 | 17.71 | 13.54 | 15.62 | 16.05 |