Has anyone here actually ran the code on their own hardware yet? I did a standard non-middleware lm_eval_harness and got 0.3214 on gpqa_main_zeroshot WITH the

Has anyone here actually ran the code on their own hardware yet? I did a standard non-middleware lm_eval_harness and got 0.3214 on gpqa_main_zeroshot WITH the systemprompt and 0.3616 without the systemprompt. Have not ran it with the middleware yet that's supposed to do the substraction. Now, if that adds 20% to the score, that would be a huge deal, but it would also roughly match the jump from gpt4o to o1-preview that they got in gpqa_diamond.