Reasoning and Question Answering on BoolQ, RTE, HellaSWAG, ARC, OpenBookQA, and PiQA
67.24Avg AccuracyBefore finetune
Evaluation Results
| Method | Links | |
|---|---|---|
| Before finetuneBackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 67.24 | |
| Target LLMBackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 66.93 | |
| ULDBackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 66.85 | |
| NPO+GDBackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 61.77 | |
| NPO+KLBackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 61.14 | |
| Offset-NPO+KLBackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 58.72 | |
| GA+GDBackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 58.34 | |
| Offset-DPO+KLBackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 56.59 | |
| DPO+KLBackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 56.34 | |
| GA+KLBackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 55.41 | |
| NPOBackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 54.73 | |
| DPO+GDBackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 53.91 | |
| Offset-GA+KLBackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 53.78 | |
| DPOBackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 48.12 | |
| GABackbone=Mistral-7B-instruct, Mode=Zero-shot2024.06 | 35.59 |