Harmlessness Evaluation on LINGUASAFE (test)
100Tie RateSHARD
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SHARDModel=Qwen3.5 9B2026.06 | 100 | 0 | 0 | |
| SHARD w/TeacherModel=Qwen3.5 27B2026.06 | 100 | 0 | 0 | |
| HHTUNEModel=Qwen3.5 9B2026.06 | 100 | 0 | 0 | |
| SHARDModel=Qwen3.5 27B2026.06 | 99.8 | 0 | 0.2 | |
| SHARD w/TeacherModel=Qwen3.5 9B2026.06 | 99.8 | 0.2 | 0 | |
| HHTUNEModel=Qwen3.5 27B2026.06 | 99.8 | 0 | 0.2 | |
| RATIONALModel=Qwen3.5 9B2026.06 | 99.8 | 0.2 | 0 | |
| RATIONALModel=Qwen3.5 27B2026.06 | 99.8 | 0 | 0.2 | |
| SHARDModel=Qwen3.5 27B2026.06 | 99.8 | — | — | |
| Benign IntentModel=Qwen3.5 27B2026.06 | 99.8 | — | — | |
| SHARDModel=Qwen3.5 9B2026.06 | 99.8 | — | — | |
| Benign IntentModel=Qwen3.5 9B2026.06 | 99.8 | — | — | |
| Benign IntentModel=Phi-42026.06 | 99.7 | — | — | |
| SHARDModel=Phi-42026.06 | 99.5 | — | — | |
| SHARD w/TeacherModel=Phi-42026.06 | 99.4 | 0.3 | 0.3 | |
| RATIONALModel=Llama 3.1 8B2026.06 | 99.4 | 0.6 | 0 | |
| SHARDModel=Phi-42026.06 | 99 | 1 | 0 | |
| HHTUNEModel=Phi-42026.06 | 99 | 0.8 | 0.2 | |
| RATIONALModel=Phi-42026.06 | 98.5 | 1.5 | 0 | |
| Benign IntentModel=Llama 3.1 8B2026.06 | 98.3 | — | — | |
| Benign IntentModel=Mistral 24B2026.06 | 98.3 | — | — | |
| SHARD w/TeacherModel=Llama 3.1 8B2026.06 | 98.2 | 0 | 1.8 | |
| HHTUNEModel=Llama 3.1 8B2026.06 | 98.2 | 1.8 | 0 | |
| SHARDModel=Llama 3.1 8B2026.06 | 97.5 | — | — | |
| SHARDModel=Mistral 24B2026.06 | 97.3 | — | — | |
| Benign IntentModel=Llama 3.3 70B2026.06 | 97.2 | — | — | |
| RATIONALModel=Mistral 24B2026.06 | 96.7 | 2.7 | 0.5 | |
| SHARDModel=Llama 3.1 8B2026.06 | 96.4 | 3.6 | 0 | |
| HHTUNEModel=Mistral 24B2026.06 | 96.4 | 3 | 0.5 | |
| SHARDModel=Llama 3.3 70B2026.06 | 96.1 | — | — | |
| SHARDModel=Mistral 7B2026.06 | 95.7 | 0.6 | 3.7 | |
| RATIONALModel=Mistral 7B2026.06 | 95.7 | 1.9 | 2.5 | |
| HHTUNEModel=Mistral 7B2026.06 | 94.4 | 2.5 | 3.1 | |
| SHARD w/TeacherModel=Mistral 24B2026.06 | 93.2 | 6.6 | 0.2 | |
| Benign IntentModel=Mistral 7B2026.06 | 92.4 | — | — | |
| SHARDModel=Mistral 24B2026.06 | 91 | 9 | 0 | |
| SHARDModel=Mistral 7B2026.06 | 90.6 | — | — | |
| SHARD w/TeacherModel=Mistral 7B2026.06 | 83.7 | 7.1 | 9.2 |