experiment/scripts/data/build_original_cache.sh

#!/usr/bin/env bash
# Build the original-model output cache from caption_targets.json.
# Run this once before running any eval script to avoid loading the base model.
#
# Usage:
#   bash experiment/scripts/data/build_original_cache.sh

SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
source "${SCRIPT_DIR}/../baselines/_common.sh"

CAPTION_TARGETS="${CAPTION_TARGETS:-${PROJECT_ROOT}/experiment/data/caption_targets.json}"
ORIGINAL_CACHE_DIR="${ORIGINAL_CACHE_DIR:-./cached_original_outputs}"

if [ ! -f "${CAPTION_TARGETS}" ]; then
    echo "ERROR: caption_targets.json not found at ${CAPTION_TARGETS}"
    echo "  Run experiment/scripts/data/run_build_caption_targets.sh first."
    exit 1
fi

echo "Building original cache..."
echo "  caption_targets: ${CAPTION_TARGETS}"
echo "  output_dir:      ${ORIGINAL_CACHE_DIR}"

python -m experiment.data.build_original_cache \
    --caption_targets "${CAPTION_TARGETS}" \
    --output_dir "${ORIGINAL_CACHE_DIR}" \
    --use_val_split