LLM Alignment on Evol-Instruct
51.4Win RateHighAvg.
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| HighAvg.Backbone=Mistral-7B, % Train=33%, Alignment Method=DPO2025.05 | 51.4 | 6.56 | |
| FullBackbone=Mistral-7B, % Train=100%, Alignment Method=DPO2025.05 | 49.5 | 6.5 | |
| HighAvg.Backbone=LLaMA-3-8B, % Train=33%, Alignment Method=SimPO2025.05 | 47.9 | 6.54 | |
| RandomBackbone=Mistral-7B, % Train=33%, Alignment Method=DPO2025.05 | 47.5 | 6.54 | |
| LowAvg.Backbone=Mistral-7B, % Train=33%, Alignment Method=DPO2025.05 | 46.8 | 6.43 | |
| FullBackbone=LLaMA-3-8B, % Train=100%, Alignment Method=DPO2025.05 | 46.1 | 6.64 | |
| FullBackbone=LLaMA-3-8B, % Train=100%, Alignment Method=SimPO2025.05 | 44.7 | 6.36 | |
| RandomBackbone=LLaMA-3-8B, % Train=33%, Alignment Method=DPO2025.05 | 42.5 | 6.13 | |
| HighAvg.Backbone=LLaMA-3-8B, % Train=33%, Alignment Method=DPO2025.05 | 42.2 | 6.28 | |
| LowAvg.Backbone=LLaMA-3-8B, % Train=33%, Alignment Method=DPO2025.05 | 39.4 | 6.13 | |
| HighAvg.Backbone=Mistral-7B, % Train=33%, Alignment Method=SimPO2025.05 | 38.5 | 5.71 | |
| HighVar.Backbone=LLaMA-3-8B, % Train=33%, Alignment Method=DPO2025.05 | 35.6 | 6.11 | |
| HighVar.Backbone=Mistral-7B, % Train=33%, Alignment Method=DPO2025.05 | 33.9 | 6.07 | |
| RandomBackbone=LLaMA-3-8B, % Train=33%, Alignment Method=SimPO2025.05 | 30.2 | 5.91 | |
| LowAvg.Backbone=LLaMA-3-8B, % Train=33%, Alignment Method=SimPO2025.05 | 30.2 | 5.93 | |
| FullBackbone=Mistral-7B, % Train=100%, Alignment Method=SimPO2025.05 | 27.5 | 5.65 | |
| HighVar.Backbone=LLaMA-3-8B, % Train=33%, Alignment Method=SimPO2025.05 | 26.1 | 6.02 | |
| ZeroshotBackbone=LLaMA-3-8B, % Train=0%, Alignment Method=DPO2025.05 | 23.4 | 5.8 | |
| ZeroshotBackbone=LLaMA-3-8B, % Train=0%, Alignment Method=SimPO2025.05 | 23.4 | 5.8 | |
| RandomBackbone=Mistral-7B, % Train=33%, Alignment Method=SimPO2025.05 | 17.7 | 5.08 | |
| HighVar.Backbone=Mistral-7B, % Train=33%, Alignment Method=SimPO2025.05 | 16.7 | 5.46 | |
| ZeroshotBackbone=Mistral-7B, % Train=0%, Alignment Method=DPO2025.05 | 16.3 | 4.87 | |
| ZeroshotBackbone=Mistral-7B, % Train=0%, Alignment Method=SimPO2025.05 | 16.3 | 4.87 | |
| LowAvg.Backbone=Mistral-7B, % Train=33%, Alignment Method=SimPO2025.05 | 16.3 | 5.1 |