Issue about reproducing results

[Ignoramus0817]

We tested MAmmoTH2-8B-Plus on Alpaca Eval 2.0 and Arena Hard, but our results are far lower than reported. Specifically,

	Alpaca Eval 2.0	Arena Hard
reported	18.5	16.6
reproduced	12.53	9.7

Note that we also tested other models like SimPO (Llama-3-Instruct-8B-SimPO-v0.2) on Alpaca Eval 2.0, the gap of results are not that large (51.54/53.7).

We use vllm-0.4.0 as inference backend and gpt-4-1106-preview as judge.
Do you have any idea where the problem is? Thanks.