eval_longbench.sh

python -m main.eval_longbench \
--data_root data \
--dataset_name qasper \
--output_dir data/outputs/debug \
--model_name_or_path ./focusllm-checkpoint \
--max_length 15500 \
--enable_beacon True \
--local_window 3500 \
--memory_stride 1024 \
--beacon_param q k v o \
--eval_data data/longbench/test.json