diff --git a/scripts/run_step1.sh b/scripts/run_step1.sh
index ce3813b..f52c81d 100755
--- a/scripts/run_step1.sh
+++ b/scripts/run_step1.sh
@@ -7,7 +7,7 @@
 # 환경 변수:
 #   DRY_RUN=1   2 epoch만 돌려 동작 검증
 #   EPOCHS=N    epoch 수 override (기본 config의 epoch_num)
-#   LOG=path    로그 파일 (기본: /workspace/PaddleOCR/output/kr_lp_pgnet/train.log)
+#   NUM_SAMPLES=N  합성 데이터 수 (기본 50000)
 
 set -euo pipefail
 
@@ -16,16 +16,22 @@ KR_LP_DIR=/workspace/kr_lp_pgnet
 TRAIN_DATA=/workspace/train_data
 SYNTH_DIR="$TRAIN_DATA/kr_lp_synth"
 ASSET_DIR="$KR_LP_DIR/data_gen/Korean-license-plate-Generator"
-LOG="${LOG:-$PADDLEOCR_DIR/output/kr_lp_pgnet/train.log}"
-NUM_SAMPLES="${NUM_SAMPLES:-10000}"
+NUM_SAMPLES="${NUM_SAMPLES:-50000}"
+
+TS=$(date +%Y%m%d_%H%M)
+RUN_NAME="step1-${TS}"
+OUTPUT_DIR="$PADDLEOCR_DIR/output/kr_lp_pgnet_${TS}"
+LOG="$OUTPUT_DIR/run.log"
 
-# ── 1. 합성 데이터 생성 ──────────────────────────────────────────────────────
 echo "==========================="
-echo "[1/3] 합성 데이터 생성 (${NUM_SAMPLES}장)"
-echo "  asset: $ASSET_DIR"
-echo "  out:   $SYNTH_DIR"
+echo "RUN: $RUN_NAME"
+echo "OUTPUT: $OUTPUT_DIR"
 echo "==========================="
 
+# ── 1. 합성 데이터 생성 ──────────────────────────────────────────────────────
+echo "[1/3] 합성 데이터 생성 (${NUM_SAMPLES}장)"
+
+rm -rf "$SYNTH_DIR"
 python3.10 "$KR_LP_DIR/data_gen/generate_synthetic.py" \
     --asset_dir "$ASSET_DIR" \
     --out_dir   "$SYNTH_DIR" \
@@ -33,29 +39,28 @@ python3.10 "$KR_LP_DIR/data_gen/generate_synthetic.py" \
     --dict      "$KR_LP_DIR/dict/kr_lp_dict.txt"
 
 # ── 2. eval GT mat 생성 ─────────────────────────────────────────────────────
-echo "==========================="
 echo "[2/3] eval GT mat 생성"
-echo "  label: $SYNTH_DIR/test/test.txt"
-echo "  out:   $SYNTH_DIR/gt/"
-echo "==========================="
 
 python3.10 "$KR_LP_DIR/tools/make_gt_mat.py" \
     --label   "$SYNTH_DIR/test/test.txt" \
     --out_dir "$SYNTH_DIR/gt"
 
 # ── 3. 학습 ─────────────────────────────────────────────────────────────────
+echo "[3/3] Step1 학습 시작"
 cd "$PADDLEOCR_DIR"
 
-# train_data symlink (config는 ./train_data/kr_lp_synth 사용)
 if [ ! -e ./train_data ]; then
     ln -sf "$TRAIN_DATA" ./train_data
 fi
 
-mkdir -p "$(dirname "$LOG")"
+mkdir -p "$OUTPUT_DIR"
 
 OVERRIDE=(
     -o Global.pretrained_model=./pretrain_models/train_step1/best_accuracy
        Global.load_static_weights=False
+       Global.save_model_dir="${OUTPUT_DIR}/"
+       Global.save_res_path="${OUTPUT_DIR}/predicts.txt"
+       wandb.name="${RUN_NAME}"
 )
 if [ -n "${EPOCHS:-}" ]; then
     OVERRIDE+=(Global.epoch_num="$EPOCHS")
@@ -65,13 +70,10 @@ if [ "${DRY_RUN:-0}" = "1" ]; then
     echo "DRY_RUN=1 → 2 epoch만 실행"
 fi
 
-echo "==========================="
-echo "[3/3] Step1 학습 시작"
-echo "  config:  configs/e2e/kr_lp_pgnet.yml"
-echo "  data:    $SYNTH_DIR/"
-echo "  pretrain: pretrain_models/train_step1/best_accuracy"
-echo "  log:     $LOG"
-echo "  override: ${OVERRIDE[@]}"
-echo "==========================="
+echo "  config:   configs/e2e/kr_lp_pgnet.yml"
+echo "  data:     $SYNTH_DIR/"
+echo "  output:   $OUTPUT_DIR/"
+echo "  wandb:    $RUN_NAME"
+echo "  log:      $LOG"
 
 python3.10 tools/train.py -c configs/e2e/kr_lp_pgnet.yml "${OVERRIDE[@]}" 2>&1 | tee "$LOG"