Long-context Understanding on LongBench-Zh filtered (test)
64.8Single Document ScoreSentinel (Qwen2.5-0.5B-Instruct)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Sentinel (Qwen2.5-0.5B-Instruct)Compression Category=Contextual Utilization Decoding, Token Constraint=2K, Proxy Model=Qwen2.5-0.5B-Instruct, Tokens=1,932, Ratio=5×, Inference Model=GPT-3.5-Turbo2025.05 | 64.8 | 25.1 | 14.3 | 34.7 | |
| Sentinel (Qwen2.5-1.5B-Instruct)Compression Category=Contextual Utilization Decoding, Token Constraint=2K, Proxy Model=Qwen2.5-1.5B-Instruct, Tokens=1,929, Ratio=5×, Inference Model=GPT-3.5-Turbo2025.05 | 63.3 | 24.9 | 14.8 | 34.3 | |
| Original PromptTokens=14,940, Inference Model=GPT-3.5-Turbo2025.05 | 61.2 | 28.7 | 16 | 35.3 | |
| LLMLingua-2Compression Category=Metric-Based Compression, Token Constraint=3K, Tokens=3,023, Ratio=5×, Inference Model=GPT-3.5-Turbo2025.05 | 46.7 | 23 | 15.3 | 28.3 | |
| LLMLinguaCompression Category=Metric-Based Compression, Token Constraint=3K, Tokens=3,060, Ratio=5×, Inference Model=GPT-3.5-Turbo2025.05 | 35.2 | 20.4 | 11.8 | 22.5 |