SFT Generalization on Tulu3 SFT (Expanded)
65.77SFT ScoreDeltaMoE
Evaluation Results
| Method | Links | |
|---|---|---|
| DeltaMoEBase Model=Qwen2.5-7B, Alignment Delta=Tulu3, Comparison Category=Baselines w/ Matched Parameters2026.05 | 65.77 | |
| LLaMA-ProBase Model=Qwen2.5-7B, Alignment Delta=Tulu3, Comparison Category=Baselines w/ Matched Parameters2026.05 | 64.97 | |
| MoLABase Model=Qwen2.5-7B, Alignment Delta=Tulu3, Comparison Category=Baselines w/ Matched Parameters2026.05 | 64.25 | |
| Dense-FT-DeltaBase Model=Qwen2.5-7B, Alignment Delta=Tulu3, Comparison Category=Baselines w/ Same CPT Data2026.05 | 62.25 | |
| Dense-FT-AvgBase Model=Qwen2.5-7B, Alignment Delta=Tulu3, Comparison Category=Baselines w/ Same CPT Data2026.05 | 61.55 | |
| Dense-FT-Delta-2FLOPsBase Model=Qwen2.5-7B, Alignment Delta=Tulu3, Comparison Category=Baselines w/ Matched FLOPs (2x CPT Data)2026.05 | 59.64 | |
| Dense-FT-Avg-2FLOPsBase Model=Qwen2.5-7B, Alignment Delta=Tulu3, Comparison Category=Baselines w/ Matched FLOPs (2x CPT Data)2026.05 | 58.32 | |
| Qwen-7B + Tulu-DeltaBase Model=Qwen2.5-7B, Alignment Delta=Tulu32026.05 | 53.85 |