Text Generation on Harry Potter forget data (400 chunks)
8.02BLEUTarget LLM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Target LLMBackbone=Mistral-7B-instruct, Status=Finetuned on forget data, Prefix length=200 tokens2024.06 | 8.02 | 16.98 | |
| NPO+GDBackbone=Mistral-7B-instruct, Prefix length=200 tokens2024.06 | 0.82 | 5.76 | |
| Before finetuneBackbone=Mistral-7B-instruct, Status=Original pretrained model, Prefix length=200 tokens2024.06 | 0.74 | 8.97 | |
| NPO+KLBackbone=Mistral-7B-instruct, Prefix length=200 tokens2024.06 | 0.74 | 6.84 | |
| ULDBackbone=Mistral-7B-instruct, Prefix length=200 tokens2024.06 | 0.67 | 4.58 | |
| Offset-NPO+KLBackbone=Mistral-7B-instruct, Prefix length=200 tokens2024.06 | 0.58 | 8.55 | |
| NPOBackbone=Mistral-7B-instruct, Prefix length=200 tokens2024.06 | 0.47 | 4.31 | |
| Offset-DPO+KLBackbone=Mistral-7B-instruct, Prefix length=200 tokens2024.06 | 0.45 | 4.39 | |
| DPO+GDBackbone=Mistral-7B-instruct, Prefix length=200 tokens2024.06 | 0.38 | 3.94 | |
| DPOBackbone=Mistral-7B-instruct, Prefix length=200 tokens2024.06 | 0.35 | 4.24 | |
| DPO+KLBackbone=Mistral-7B-instruct, Prefix length=200 tokens2024.06 | 0.35 | 4.15 | |
| GABackbone=Mistral-7B-instruct, Prefix length=200 tokens2024.06 | 0 | 0 | |
| GA+GDBackbone=Mistral-7B-instruct, Prefix length=200 tokens2024.06 | 0 | 0 | |
| GA+KLBackbone=Mistral-7B-instruct, Prefix length=200 tokens2024.06 | 0 | 0 | |
| Offset-GA+KLBackbone=Mistral-7B-instruct, Prefix length=200 tokens2024.06 | 0 | 0 |