mxnv_acc.md

Average accuracy of hellaswag,lambada_openai,mmlu,piqa,winogrande.

We evaluated using a fake model since we currently have no access to devices for running the real models. However, we have verified that in most cases the fake model closely matches the real model.

mxfp4 g32	llama3.1-8B-Instruct	Qwen2-7.5-Instruct	Phi4	Qwen3-32B
RTN	0.6212	0.6550	0.7167	0.6901
AutoRound	0.6686	0.6758	0.7247	0.7211
AutoRound+alg_ext	0.6732	0.6809	0.7225	0.7201

nvfp4 g16	llama3.1-8B-Instruct	Qwen2-7.5-Instruct	Phi4	Qwen3-32B
RTN	0.6876	0.6906	0.7296	0.7164
AutoRound	0.6918	0.6973	0.7306	0.7306
AutoRound+alg_ext	0.6965	0.6989	0.7318	0.7295