A set of 50% weight-sparse Llama3.1-8B pruned by Wanda. Model links are in the table below. Models can be loaded as is with Huggingface Transformers.
Perplexity
MMLU (5-shot)
| MMLU (5-shot) | Accuracy (%) | Relative to Dense (%) | Model Link |
|---|---|---|---|
| Dense | 65.1 | baseline | Meta-Llama-3.1-8B-wanda-unstructured-0.0 |
| Unstructured | 50.0 | -15.1 | Meta-Llama-3.1-8B-wanda-unstructured-0.5 |
| 4:8 | 39.3 | -25.8 | Meta-Llama-3.1-8B-wanda-4of8 |
| 2:4 | 28.7 | -36.4 | Meta-Llama-3.1-8B-wanda-2of4 |