Can't reproduce the result

[GGgrey]

I use the base alignment-handbook/zephyr-7b-sft-full model, revision: ac6e600eefcce74f5e8bae1035d4f66019e93190, lm-evaluation-harness version v0.4.0, but i can't reproduce the result as paper on the hellaswag task, can you help me?

[GGgrey]

The same goes for gsm8k, although the results are very close to the paper.