appvoid/arco-2

This is a passthrough of arco with an experimental model. It improved on arc challenge, only missing 1.2 points to get to the level of modern 3b baseline performance.

If you prefer answering multilingual, general knowledge, trivially simple questions chose qwen or llama. If you prefer solving trivially simple english tasks while being half the size, chose arco.

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	qwen 2	44.13	28.92	49.05	69.31	56.99	49.68
0.3b	smollm2	25.52	37.71	56.41	71.93	59.27	50.17
0.5b	qwen 2.5	47.29	31.83	52.17	70.29	57.06	51.72
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
1.24b	llama 3.2	36.75	36.18	63.70	74.54	60.54	54.34

supporters

trivia

arco also means "arc optimized" hence the focus on this cognitive-based benchmark.