Reproduce Llama 2 results

[EttoreCaputo]

Hi, I'm trying to reproduce the Llama-2-7B results in your paper (ACC and F1 in Zero Shot) but the values are very different. How did you execute the tasks with Llama 2? In my case I only wrote this:

!python main.py \
    --model_name llama2_7b \
    --path_model meta-llama/Llama-2-7b-hf \
    --task {task_type} \
    --data_name {dataset_name} \
    --num_train {num_train}

in a Kaggle Notebook and used the 2 x Tesla T4 accelerator.
For example I'm getting 18.04 (F1 score) on IAM claims and not 60.14. What I'm doing wrong?

P.S. I changed also the tensor_parallel_size arg in modeling.py from 1 to 2.