#!/bin/bash
# Eval AB2 (w/o ST-Aug) ckpt956 — mirrors run_eval_v10_r2_ckpt956.sh.
set -e

cd /mnt/local-fast/zhangt/forensics_grpo

MODEL=outputs_forensics/ab_noAug/checkpoint-956
OUT=eval_ab_noAug_ckpt956
mkdir -p "$OUT/logs"

export PATH="/mnt/local-fast/zhangt/torch_env/bin:$PATH"
export LD_LIBRARY_PATH="/opt/conda/lib:${LD_LIBRARY_PATH}"
export PYTHONPATH=".:$PYTHONPATH"

for R in 0 1 2 3 4 5 6 7; do
  CUDA_VISIBLE_DEVICES=$R python evaluate_forensics.py \
      --model_path "$MODEL" \
      --rank $R --world_size 8 --device 0 \
      --out_dir "$OUT" \
      --cot false --max_new_tokens 64 --temperature 0.0 \
      > "$OUT/logs/rank_${R}.log" 2>&1 &
done
wait
echo "all 8 ranks done"

python evaluate_grounding_metrics.py --out_dir "$OUT" | tee "$OUT/grounding_metrics.txt"