appvoid/palmer-004

june update

This model has improved overall performance at the expense of small degradation on winogrande. As all palmer models, the model is biased to respond to answers without using any specific prompt, feel free to further fine-tune it for your specific use case.

Model	MMLU	ARC-C	HellaSwag	PIQA	Winogrande	Average
tinyllama-3t	0.2577	0.3029	0.5935	0.7329	0.5959	0.4966
palmer-004-old	0.2601	0.3456	0.6138	0.7443	0.6511	0.5229
palmer-004	0.2661	0.3490	0.6173	0.7481	0.6417	0.5244

note

Even though palmer-003 is only 2k context size, it's 0.5257 on average so if you don't need that much context size (32k) you are better off with the former one.