download "gauntlet" for summarization (peter's version) and run summarization inference on it with the textsum package

dl_gauntlet.sh

URL=https://www.dropbox.com/sh/zu1p7rhg5238a5y/AABsJN_pCYf9plSDZY8ziKATa?dl=1
wget -O docs.zip $URL
unzip -B -j docs.zip -d gauntlet && rm -rf docs.zip

run_gauntlet.sh

#!/bin/bash
# run pip install textsum or pip install textsum[8bit]

# Set-able options
NUM_BEAMS=4
BATCH_SIZE=16384
MODEL="pszemraj/long-t5-tglobal-base-16384-book-summary"

# Extract the base name of the model
MODEL_BASENAME=$(basename "${MODEL}")

# Input directory
INPUT_DIR="gauntlet/"

# Output directory
OUTPUT_DIR="${MODEL_BASENAME}_output"

# Create output directory if it doesn't exist
mkdir -p "${OUTPUT_DIR}"

# Run the textsum-dir command with the specified options
# add -8bit if you want to do LLM.int8 quantization with bitsandbytes (requires install with [8bit]
# add --tf32 if for TF32 computation on ampere gpu or newer

textsum-dir --shuffle -m "${MODEL}" \
    -batch "${BATCH_SIZE}" -nb "${NUM_BEAMS}" \
    "${INPUT_DIR}" -o "${OUTPUT_DIR}"

# Compress the output directory into a semantically-named zip file
ZIP_FILE="${MODEL_BASENAME}_gauntlet_summary-bs-${BATCH_SIZE}-nb-${NUM_BEAMS}.zip"
zip -r -9 -j "${ZIP_FILE}" "${OUTPUT_DIR}"

echo "Script finished. The output is compressed to ${ZIP_FILE}."