Long-context Reasoning on LongBench and NIAH (test)
29.33MultiFieldQA ScoreSmolLM-DroPE
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| SmolLM-DroPEZero-shot length generalization=true, RoPE scaling extension=DroPE2025.12 | 29.33 | 7.93 | 21.87 | 18.56 | 74.92 | 30.52 | |
| SmolLM + YaRNZero-shot length generalization=true, RoPE scaling extension=YaRN2025.12 | 20.78 | 4.77 | 15.03 | 10.87 | 48.25 | 19.94 | |
| SmolLM + RoPE-NTKZero-shot length generalization=true, RoPE scaling extension=RoPE-NTK2025.12 | 18.87 | 4.89 | 23.71 | 8.26 | 29.84 | 17.11 | |
| SmolLM + PIZero-shot length generalization=true, RoPE scaling extension=PI2025.12 | 13.68 | 2.45 | 5.67 | 11.52 | 0 | 6.66 | |
| SmolLMZero-shot length generalization=true2025.12 | 4.03 | 0.4 | 4.48 | 5.99 | 0 | 2.98 |