Language Modeling on OpenR1-Math seed-1 representative
2.86Perplexity (PPL)CE + TMKL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CE + TMKLAdapter=DoRA2026.05 | 2.86 | 0.35 | |
| CE + TMKLAdapter=LoRA2026.05 | 2.87 | 0.34 | |
| CE + TMKLAdapter=SineLoRA2026.05 | 2.87 | 0.34 | |
| CEAdapter=RandLoRA2026.05 | 2.88 | 0.33 | |
| CE + TMKLAdapter=RandLoRA2026.05 | 2.92 | 0.29 | |
| CEAdapter=SineLoRA2026.05 | 3 | 0.21 | |
| CEAdapter=LoRA2026.05 | 3.05 | 0.16 | |
| CEAdapter=DoRA2026.05 | 3.05 | 0.16 | |
| no adaptationAdapter=Base Qwen2.5-0.5B2026.05 | 3.21 | — |