metadata
license: apache-2.0
Zero-shot results when using the Llama-3.1-70B-Instruct as the teacher model, and the Llama-3.2-3B-Instruct as the initialized model
| Task | Llama-3.2-3B-Instruct | Llama3.2-Mamba-3B-distill |
|---|---|---|
| arc_challenge | 0.459 | 0.4838 |
| arc_easy | 0.7407 | 0.7765 |
| hellaswag | 0.7043 | 0.7037 |
| mmlu | 0.6043 | 0.5448 |
| openbookqa | 0.36 | 0.394 |
| piqa | 0.7568 | 0.7731 |
| pubmedqa | 0.696 | 0.664 |
| race | 0.4067 | 0.4029 |
| winogrande | 0.6748 | 0.6732 |
@article{junxiongdaniele2024mambainllama,
title = {The Mamba in the Llama: Distilling and Accelerating Hybrid Models},
author = {Junxiong Wang and Daniele Paliotta and Avner May and Alexander M. Rush and Tri Dao},
journal = {arXiv preprint arXiv:2408.15237},
year = {2024}
}