appvoid/arco-2.1

This is the latest iteration as an effort to make arco as good on arc as it can. So far it improved a little.

prompt

there is no prompt intentionally set.

benchmarks

zero-shot results from state-of-the-art small language models

Parameters	Model	MMLU	ARC-C	HellaSwag	PIQA	Winogrande	Average
0.5b	danube 3	24.81	36.18	60.46	73.78	61.01	51.25
0.5b	arco	26.17	37.29	62.88	74.37	62.27	52.60
0.5b	arco 2	25.51	38.82	63.02	74.70	61.25	52.66
0.5b	arco 2º	25.47	38.99	63.03	74.70	61.01	52.64

supporters

trivia

arco seems to keep improving on the same 3 benchmarks, reached its limit though.