something-else/9BQuetzal-rocm-HF-1

This 9B model, built on the RWKV v5 architecture, was exclusively trained using AMD GPUs. The model's training process advanced in tandem with the evolution of ROCm (upto ROCm 6.0.0), this means a lot of experimentation 😅.

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
mathqa	Yaml	none	0	acc	0.2673	±	0.0081
		none	0	acc_norm	0.2747	±	0.0082
copa	Yaml	none	0	acc	0.87	±	0.0338
boolq	Yaml	none	0	acc	0.6927	±	0.0081
hellaswag	Yaml	none	0	acc	0.5148	±	0.0050
		none	0	acc_norm	0.6833	±	0.0046
sciq	Yaml	none	0	acc	0.9430	±	0.0073
		none	0	acc_norm	0.9210	±	0.0085
lambada_openai	Yaml	none	0	perplexity	3.7234	±	0.0767
		none	0	acc	0.7145	±	0.0063
piqa	Yaml	none	0	acc	0.7568	±	0.0100
		none	0	acc_norm	0.7693	±	0.0098
arc_challenge	Yaml	none	0	acc	0.3823	±	0.0142
		none	0	acc_norm	0.4172	±	0.0144
arc_easy	Yaml	none	0	acc	0.7151	±	0.0093
		none	0	acc_norm	0.7109	±	0.0093