General capability on OBQA (test)
44Normalized AccuracyDepth Pruning
Evaluation Results
| Method | Links | |
|---|---|---|
| Depth PruningRatio=25%, Recovery finetuning=w/, Base Model=Llama-3.1-8B-Instruct2026.06 | 44 | |
| Llama-3.1-8B-InstructRatio=0%, Base Model=Llama-3.1-8B-Instruct2026.06 | 43 | |
| Adapt-PrunerRatio=25%, Recovery finetuning=w/o, Base Model=Llama-3.1-8B-Instruct2026.06 | 41 | |
| Adapt-PrunerRatio=25%, Recovery finetuning=w/, Base Model=Llama-3.1-8B-Instruct2026.06 | 40 | |
| Persona-PrunerRatio=25%, Recovery finetuning=w/, Base Model=Llama-3.1-8B-Instruct2026.06 | 40 | |
| Depth PruningRatio=25%, Recovery finetuning=w/o, Base Model=Llama-3.1-8B-Instruct2026.06 | 39 | |
| LLM-PrunerRatio=25%, Recovery finetuning=w/, Base Model=Llama-3.1-8B-Instruct2026.06 | 39 | |
| Persona-PrunerRatio=25%, Recovery finetuning=w/o, Base Model=Llama-3.1-8B-Instruct2026.06 | 39 | |
| Depth PruningRatio=25%, Recovery finetuning=w/, Base Model=Llama-3.2-3B-Instruct2026.06 | 38 | |
| SliceGPTRatio=25%, Recovery finetuning=w/, Base Model=Llama-3.1-8B-Instruct2026.06 | 38 | |
| LLM-PrunerRatio=25%, Recovery finetuning=w/, Base Model=Llama-3.2-3B-Instruct2026.06 | 37 | |
| LLM-PrunerRatio=25%, Recovery finetuning=w/o, Base Model=Llama-3.1-8B-Instruct2026.06 | 37 | |
| Depth PruningRatio=50%, Recovery finetuning=w/, Base Model=Llama-3.1-8B-Instruct2026.06 | 37 | |
| Llama-3.2-3B-InstructRatio=0%, Base Model=Llama-3.2-3B-Instruct2026.06 | 36 | |
| Adapt-PrunerRatio=25%, Recovery finetuning=w/o, Base Model=Llama-3.2-3B-Instruct2026.06 | 36 | |
| Adapt-PrunerRatio=25%, Recovery finetuning=w/, Base Model=Llama-3.2-3B-Instruct2026.06 | 36 | |
| Persona-PrunerRatio=25%, Recovery finetuning=w/o, Base Model=Llama-3.2-3B-Instruct2026.06 | 36 | |
| Persona-PrunerRatio=25%, Recovery finetuning=w/, Base Model=Llama-3.2-3B-Instruct2026.06 | 36 | |
| Persona-PrunerRatio=50%, Recovery finetuning=w/, Base Model=Llama-3.1-8B-Instruct2026.06 | 36 | |
| LLM-PrunerRatio=25%, Recovery finetuning=w/o, Base Model=Llama-3.2-3B-Instruct2026.06 | 35 | |
| LLM-PrunerRatio=50%, Recovery finetuning=w/, Base Model=Llama-3.1-8B-Instruct2026.06 | 35 | |
| Adapt-PrunerRatio=50%, Recovery finetuning=w/, Base Model=Llama-3.1-8B-Instruct2026.06 | 35 | |
| LLM-PrunerRatio=50%, Recovery finetuning=w/, Base Model=Llama-3.2-3B-Instruct2026.06 | 34 | |
| Persona-PrunerRatio=50%, Recovery finetuning=w/, Base Model=Llama-3.2-3B-Instruct2026.06 | 34 | |
| SliceGPTRatio=25%, Recovery finetuning=w/o, Base Model=Llama-3.1-8B-Instruct2026.06 | 34 | |
| Persona-PrunerRatio=50%, Recovery finetuning=w/o, Base Model=Llama-3.1-8B-Instruct2026.06 | 34 | |
| Depth PruningRatio=25%, Recovery finetuning=w/o, Base Model=Llama-3.2-3B-Instruct2026.06 | 33 | |
| Adapt-PrunerRatio=50%, Recovery finetuning=w/, Base Model=Llama-3.2-3B-Instruct2026.06 | 33 | |
| SliceGPTRatio=25%, Recovery finetuning=w/, Base Model=Llama-3.2-3B-Instruct2026.06 | 32 | |
| Depth PruningRatio=50%, Recovery finetuning=w/, Base Model=Llama-3.2-3B-Instruct2026.06 | 32 | |
| Persona-PrunerRatio=50%, Recovery finetuning=w/o, Base Model=Llama-3.2-3B-Instruct2026.06 | 32 | |
| LLM-PrunerRatio=50%, Recovery finetuning=w/o, Base Model=Llama-3.1-8B-Instruct2026.06 | 32 | |
| SliceGPTRatio=25%, Recovery finetuning=w/o, Base Model=Llama-3.2-3B-Instruct2026.06 | 31 | |
| Depth PruningRatio=50%, Recovery finetuning=w/o, Base Model=Llama-3.1-8B-Instruct2026.06 | 31 | |
| LLM-PrunerRatio=50%, Recovery finetuning=w/o, Base Model=Llama-3.2-3B-Instruct2026.06 | 30 | |
| Adapt-PrunerRatio=50%, Recovery finetuning=w/o, Base Model=Llama-3.2-3B-Instruct2026.06 | 30 | |
| Adapt-PrunerRatio=50%, Recovery finetuning=w/o, Base Model=Llama-3.1-8B-Instruct2026.06 | 30 | |
| SliceGPTRatio=50%, Recovery finetuning=w/o, Base Model=Llama-3.1-8B-Instruct2026.06 | 28 | |
| SliceGPTRatio=50%, Recovery finetuning=w/, Base Model=Llama-3.1-8B-Instruct2026.06 | 28 | |
| Depth PruningRatio=50%, Recovery finetuning=w/o, Base Model=Llama-3.2-3B-Instruct2026.06 | 26 | |
| SliceGPTRatio=50%, Recovery finetuning=w/o, Base Model=Llama-3.2-3B-Instruct2026.06 | 26 | |
| SliceGPTRatio=50%, Recovery finetuning=w/, Base Model=Llama-3.2-3B-Instruct2026.06 | 24 |