JunxiongWang/Mamba2InLlama3B_Half_DPO

Zero-shot results when using the Llama-3.1-8B-Instruct as the teacher model, and the Llama-3.2-3B-Instruct as the initialized model

Model	Llama-3.2-3B-Instruct	Llama-3.2-Mamba2-0.5-3B-sft	Llama-3.2-Mamba2-0.5-3B-dpo
Initialization Model	N/A	Llama-3.2-3B-Instruct	Llama-3.2-3B-Instruct
Teacher Model	N/A	Llama-3.1-8B-Instruct	Llama-3.1-8B-Instruct
arc_challenge	0.459	0.4667	0.541
arc_easy	0.7407	0.7668	0.8026
hellaswag	0.7043	0.6913	0.7445
mmlu	0.6043	0.5271	0.5247
openbookqa	0.36	0.388	0.424
piqa	0.7568	0.7601	0.7769
pubmedqa	0.696	0.638	0.654
race	0.4067	0.3981	0.4344
winogrande	0.6748	0.6606	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}
}