[Question]: How can I reproduce the FullAttention results on the Ruler dataset

Describe the issue

Thank you for such a solid piece of work. I followed the instructions in the README and ran the following command:
bash run.sh gradientai/Llama-3-8B-Instruct-262k hf ./result

However, I couldn't reproduce the 4K context length results mentioned in the paper.
The results are as shown in the attached image.
Am I doing something wrong?

 0              1              2              3                4                5                6                7                8      9     10     11    12    13

0 Tasks niah_single_1 niah_single_2 niah_single_3 niah_multikey_1 niah_multikey_2 niah_multikey_3 niah_multivalue niah_multiquery vt cwe fwe qa_1 qa_2
1 Score 100.0 100.0 100.0 100.0 100.0 100.0 83.0 100.0 100.0 94.4 93.33 64.0 44.0
2 Nulls 0/25 0/25 0/25 0/25 0/25 0/25 0/25 0/25 0/25 0/25 0/25 0/25 0/25