Information Retrieval on NIAH single v1.0 (test)
100AccuracyOriginal
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| OriginalBackbone=Qwen3-1.7B-Base, Context Length=4k, SWA ratio=0.752026.04 | 100 | — | |
| BOSCHBackbone=Qwen3-14B-Base, Context Length=32k, Training Tokens=2.5B, SWA ratio=0.752026.04 | 100 | — | |
| BOSCHBackbone=Qwen3-8B-Base, Context Length=32k, Training Tokens=2.5B, SWA ratio=0.752026.04 | 99.9 | — | |
| B-layerBackbone=Qwen3-1.7B-Base, Context Length=4k, Training Tokens=0.7B, SWA ratio=0.752026.04 | 99.7 | — | |
| BOSCHBackbone=Qwen3-30B-A3B-Base, Context Length=32k, Training Tokens=2.5B, SWA ratio=0.752026.04 | 99.6 | — | |
| FISHERBackbone=Qwen3-1.7B-Base, Context Length=4k, Training Tokens=0.7B, SWA ratio=0.752026.04 | 99.5 | — | |
| BOSCHBackbone=Qwen3-1.7B-Base, Context Length=4k, Training Tokens=0.7B, SWA ratio=0.752026.04 | 98.4 | — | |
| INTRBackbone=Qwen3-1.7B-Base, Context Length=4k, Training Tokens=0.7B, SWA ratio=0.752026.04 | 95.9 | — |