lm_eval --model hf \
    --model_args pretrained=/share/pretrained_lm/Qwen/Qwen2-0.5B\
    --tasks "ja/jaqket_v1" \
    --device cuda:0 \
    --batch_size 8