Question about the reproduction of COPA results

[hyeseongshin]

Hi, thanks for your great works!
I have some questions about the reproduction of COPA results.

First of all, please understand that i use translator as a problem of my english writing skills.

I tried to reproduce the evaluation results of COPA 0-shot using the LLaMA-7B model in figure-8 of the paper.
I use the script in "h2o_hf/scripts/lm_eval" and also use "huggyllama/LLaMA-7b" model.

As shown in the graph below, Full and Local were reproduced similarly, but Heavy-Hitter Oracle showed slightly different results.

-my result-

-figure-8 of the paper-

This graph is when heavy_ratio and recent_ratio are set equally. For example, when KV Cache Budget is 60%, heavy_ratio and recent_ratio are set equal to 0.3.
Different result came out, so I experimented with recent_ratio always being 0.1. (except KV Cache Budget is 10%)
I didn't draw a graph, but likewise I couldn't reproduce the results.

Can't figure out the reason about it. Would you please give me some advice?
Thanks a lot!