uv run python docgenie/evaluation/runners/per_sample_evaluator.py \
    --dataset-name rvlcdip \
    --model-name microsoft/layoutlmv3-base \
    --run-name train-exp-rvlcdip-with-ssl-layoutlmv3-v2 \
    --monitored-metric validation/f1 \
    --eval-batch-size 1 \
    --use-segment-level-bboxes \
    --tokenizer-name microsoft/layoutlmv3-base \
    --split-for-per-sample-eval test \
    --no-do-train \
    $@