Long-context Understanding on LongBench 32k
41.35LongBench 32k ScoreGemma-3-4B
Evaluation Results
| Method | Links | |
|---|---|---|
| Gemma-3-4B2026.04 | 41.35 | |
| Qwen3-4BVersion=25072026.04 | 40 | |
| SpB2.0-5BSFT Strategy=General SFT2026.04 | 37.04 | |
| Llama3.2-3B2026.04 | 33.73 | |
| Qwen2.5-3B2026.04 | 28.22 | |
| SpB1.0-7B2026.04 | 27.5 | |
| SWA-128-NoPEScale=S5 (0.66B), Training Tokens=≈ 100B2026.06 | 19.46 | |
| FullScale=S5 (0.66B), Training Tokens=≈ 100B2026.06 | 18.93 | |
| SWA-128Scale=S5 (0.66B), Training Tokens=≈ 100B2026.06 | 18.3 |