evaluation_code/scripts/eval_ruler_bench.sh

#!/usr/bin/env bash
set -euo pipefail
REPO_ROOT=$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)
export PYTHONPATH="$REPO_ROOT/src${PYTHONPATH:+:$PYTHONPATH}"

if [[ $# -lt 3 || $# -gt 4 ]]; then
  echo "Usage: $0 RULER_METADATA_JSON VIDEOS_DIR PREPARED_DIR [SEEDS]" >&2
  exit 2
fi

METADATA_JSON=$1
VIDEOS_DIR=$2
PREPARED_DIR=$3
SEEDS=${4:-0}

python -m vwg_bench.cli prepare-ruler \
  --metadata-json "$METADATA_JSON" \
  --videos-dir "$VIDEOS_DIR" \
  --output-dir "$PREPARED_DIR" \
  --seeds "$SEEDS"

echo "Prepared RULER-Bench inputs at: $PREPARED_DIR"
echo "Run the pinned upstream RULER-Bench evaluator separately; see external_benchmarks/README.md."