build_remaining_forgets.sh

#!/usr/bin/env bash
# Parallel ToMBench forget-set build for the remaining 21 topics (all except
# social_life, science_math_and_technology, adult_content which are already
# built). 7 jobs x 3 topics each; each scans the shards once per topic (~7min).
set -uo pipefail
REPO="$HOME/dev/data-attribution-comma-attr"
TDA=/storage/ice-shared/cs7634/staff/TDA
PDP=$TDA/trackstar/scores_fused/comma_2t/comma_2t_holdout/aggregates/per_doc/comma_2t_queries_olmes_tombench_per_doc_influence.parquet
OUT=$TDA/comma/forget_sets/comma_2t/tombench
cd "$REPO"
# NOTE: do NOT name this array GROUPS — that is a readonly bash builtin (user GIDs).
TGROUPS=(
    "art_and_design crime_and_law education_and_jobs"
    "electronics_and_hardware entertainment fashion_and_beauty"
    "finance_and_business food_and_dining games"
    "health history_and_geography home_and_hobbies"
    "industrial literature politics"
    "religion software software_development"
    "sports_and_fitness transportation travel_and_tourism"
)
i=0
for g in "${TGROUPS[@]}"; do
    i=$((i + 1))
    jid=$(sbatch --parsable --job-name="c2ttom_fb_${i}" \
        --export="ALL,PER_DOC_PARQUET=${PDP},OUTPUT_DIR=${OUT},PROBE=tombench,N_DOCS=200,MIN_TOKENS=0,TOPICS=${g}" \
        scripts/slurm/unlearning/comma_build_forgetsets.sbatch)
    echo "c2ttom_fb_${i} [$g] -> $jid"
done