ARC accuracy metric Reproduction

[jiaweif3ng]

Hi. Akari!
I cannot reproduce your results on the arc-c dataset.
The model I used is Llama2-7B. My result under Baselines without retrieval is 26.38%, and the result under Baselines with retrieval is 20.29%, which is too different from the results in the paper. The retrieval documents I used are the wikitext you provided.