launch_expand_train.sh

#!/usr/bin/env bash
# Launch the 66 additional Comma-2T ToMBench expA cells (22 topics x 3 seeds),
# train-only, identical settings to launch_tom_train.sh but the expand manifest.
set -uo pipefail
REPO_DIR="${REPO_DIR:-$HOME/dev/data-attribution-comma-attr}"
cd "$REPO_DIR"
source scripts/comma/comma_2t_sweep_env.sh
export EXTRA_OVERRIDES="$SWEEP_PAPER_OVERRIDES"
export MAXWALL=340
TDA=/storage/ice-shared/cs7634/staff/TDA
SUBMIT_FLAG=""
[ "${1:-}" = "--submit" ] && SUBMIT_FLAG="--submit"
./.venv/bin/python scripts/unlearning/submit_faithful_unlearning.py \
    --manifest "$TDA/comma/tom_unlearning/manifest_expand.tsv" \
    --runs-root "$TDA/comma/tom_unlearning/runs" \
    --dolma-cache "$TDA/comma/retain_cache" \
    --model-config comma_2t_lora \
    --max-forget-docs 200 \
    --batch-size 4 --grad-accum 4 \
    --log-dir "$SWEEP_LOG_DIR" \
    --job-prefix c2ttomtr \
    --exclude "$SWEEP_EXCLUDE_NODES" \
    --constraint "$SWEEP_GPU_CONSTRAINT" \
    $SUBMIT_FLAG