_runs/launch_strawman_cellpolicy.sh

#!/usr/bin/env bash
# Launch two strawman variants (single-stage cell-policy at stage_i=3, no
# curriculum, no thought tokens) on GPUs 0 and 1.
set -euo pipefail
ROOT="${ROOT:-/home/ubuntu/curriculum_cot}"
TS="$(date +%Y%m%d_%H%M%S)"
SWEEP_ROOT="${ROOT}/_runs/strawman_cellpolicy_${TS}"
mkdir -p "${SWEEP_ROOT}"
PIPE="${ROOT}/_runs/strawman_cellpolicy_pipeline.sh"
chmod +x "${PIPE}"

launch() {
  # Usage: launch <variant> <gpu> <KEY=VALUE>...
  local variant="$1" gpu="$2"
  shift 2
  local out="${SWEEP_ROOT}/${variant}"
  mkdir -p "${out}"
  echo "[launch] ${variant} on GPU ${gpu}  out=${out}"
  nohup env VARIANT="${variant}" GPU="${gpu}" OUTPUT_ROOT="${out}" "$@" \
       bash "${PIPE}" > "${out}/console.log" 2>&1 &
  local pid=$!
  disown "${pid}" || true
  echo "${variant}=${pid}" >> "${SWEEP_ROOT}/PIDS.txt"
}

launch strawman_a_lr2e5 0 \
  SFT_LR=2e-5 GRPO_LR=5e-6 SFT_MAX_STEPS=3000 GRPO_MAX_STEPS=1500 \
  PENALTY_MISSING=1.0 EXACT_MATCH_BONUS=1.0 CARD_MISMATCH_PEN=1.5 \
  SFT_OVERSAMPLE=3

launch strawman_b_lr5e5 1 \
  SFT_LR=5e-5 GRPO_LR=5e-6 SFT_MAX_STEPS=4000 GRPO_MAX_STEPS=1500 \
  PENALTY_MISSING=1.0 EXACT_MATCH_BONUS=1.0 CARD_MISMATCH_PEN=1.5 \
  SFT_OVERSAMPLE=3

echo "[launch] sweep root: ${SWEEP_ROOT}"
echo "[launch] PIDs:"
cat "${SWEEP_ROOT}/PIDS.txt"