evaluation_code/scripts/eval_v_reasonbench.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 4 || $# -gt 5 ]]; then
  echo "Usage: $0 VREASON_METADATA_JSON VIDEOS_DIR PREPARED_DIR MODEL_NAME [SEEDS]" >&2
  exit 2
fi

METADATA_JSON=$1
VIDEOS_DIR=$2
PREPARED_DIR=$3
MODEL_NAME=$4
SEEDS=${5:-0,1,2,3,4}

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

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