Long-context language understanding on LongBench En
38.84Single-Document ScoreOriginal Prompt
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Original PromptDownstream LLM=Qwen2.5-Instruct-7B2025.05 | 38.84 | 44.74 | 22.76 | 35.45 | 37.3 | |
| SentinelProxy Model=Qwen2.5-0.5B, Context Constraint=2K, Downstream LLM=Qwen2.5-Instruct-7B2025.05 | 37.73 | 46.16 | 23.03 | 35.64 | 38.02 | |
| Raw AttentionProxy Model=Qwen2.5-0.5B, Context Constraint=2K, Downstream LLM=Qwen2.5-Instruct-7B2025.05 | 34.92 | 38.96 | 21.32 | 31.74 | 33.12 | |
| RandomContext Constraint=2K, Downstream LLM=Qwen2.5-Instruct-7B2025.05 | 28.22 | 30.68 | 20.33 | 26.41 | 28.3 | |
| EmptyContext Constraint=2K, Downstream LLM=Qwen2.5-Instruct-7B2025.05 | 10.72 | 22.26 | 16.46 | 16.48 | 16.05 |