Value Alignment on BeaverTails (test)
59.9Value Alignment ScoreNo Guidance
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| No GuidanceBackbone=QWEN2-1.5B2026.05 | 59.9 | — | |
| No GuidanceBackbone=LLAMA-3.2-3B2026.05 | 58.95 | — | |
| No GuidanceBackbone=MISTRAL-7B2026.05 | 50.9 | — | |
| System PromptBackbone=LLAMA-3.2-3B2026.05 | 42.04 | 28.7 | |
| System PromptBackbone=MISTRAL-7B2026.05 | 41.4 | 18.7 | |
| ITIBackbone=LLAMA-3.2-3B2026.05 | 40.63 | 31.1 | |
| RE-ControlBackbone=LLAMA-3.2-3B2026.05 | 39.27 | 33.4 | |
| ITIBackbone=MISTRAL-7B2026.05 | 37.2 | 26.9 | |
| No GuidanceBackbone=GPT-22026.05 | 35.94 | — | |
| DPO (LoRA)Backbone=LLAMA-3.2-3B2026.05 | 34.71 | 41.1 | |
| System PromptBackbone=QWEN2-1.5B2026.05 | 34.49 | 42.4 | |
| RE-ControlBackbone=QWEN2-1.5B2026.05 | 34.25 | 42.8 | |
| ITIBackbone=QWEN2-1.5B2026.05 | 33.84 | 43.5 | |
| System PromptBackbone=GPT-22026.05 | 33.48 | 6.8 | |
| RE-ControlBackbone=MISTRAL-7B2026.05 | 32.12 | 36.9 | |
| DPO (LoRA)Backbone=QWEN2-1.5B2026.05 | 30.56 | 49 | |
| ITIBackbone=GPT-22026.05 | 28.85 | 19.7 | |
| SVGTBackbone=LLAMA-3.2-3B2026.05 | 28.58 | 51.5 | |
| SVGTBackbone=QWEN2-1.5B2026.05 | 28.04 | 53.2 | |
| DPO (LoRA)Backbone=MISTRAL-7B2026.05 | 24.16 | 52.5 | |
| RE-ControlBackbone=GPT-22026.05 | 23.07 | 35.8 | |
| DPO (LoRA)Backbone=GPT-22026.05 | 20.64 | 42.6 | |
| SVGTBackbone=GPT-22026.05 | 19.89 | 44.7 | |
| SVGTBackbone=MISTRAL-7B2026.05 | 13.4 | 73.7 |