Helpfulness Evaluation on MTBench
9.35HelpfulnessGPT-4o
Evaluation Results
| Method | Links | |
|---|---|---|
| GPT-4oNBF steering=false2025.02 | 9.35 | |
| o1NBF steering=false2025.02 | 9.22 | |
| Claude 3.5 SonnetNBF steering=false2025.02 | 9.14 | |
| o1NBF steering=true2025.02 | 8.83 | |
| GPT-4oNBF steering=true2025.02 | 8.77 | |
| Claude 3.5 SonnetNBF steering=true2025.02 | 8.61 | |
| GPT-3.5-turboNBF steering=false2025.02 | 8 | |
| GPT-3.5-turboNBF steering=true2025.02 | 7.59 | |
| Llama3.1-8B-InstructFL method=FedAvg2025.02 | 6.8 | |
| Llama3.1-8B-InstructFL method=SCAFFOLD2025.02 | 6.8 | |
| FL + Safety filter + CAIFL method=FedAvg2025.02 | 6.1 | |
| FL + CAIFL method=SCAFFOLD2025.02 | 5.9 | |
| FL + CAIFL method=FedAvg2025.02 | 5.8 | |
| FL + Safety filter + CAIFL method=SCAFFOLD2025.02 | 5.8 | |
| FLFL method=SCAFFOLD2025.02 | 2.9 | |
| FLFL method=FedAvg2025.02 | 2.7 | |
| FL + Safety filterFL method=SCAFFOLD2025.02 | 2.7 | |
| FL + Safety filterFL method=FedAvg2025.02 | 2.4 |