diff --git a/sanguo_factor/batch_eval.py b/sanguo_factor/batch_eval.py
index d522c94..c0815ab 100644
--- a/sanguo_factor/batch_eval.py
+++ b/sanguo_factor/batch_eval.py
@@ -62,7 +62,7 @@ def run_batch_eval(
         raise ValueError(f"股票池为空: vnpy_db={vnpy_db} window={start}~{end}")
 
     alpha_df = bars.select(["vt_symbol", "datetime", "open", "high", "low", "close",
-                            "volume", "turnover", "open_interest", "vwap"])
+                            "volume", "turnover", "vwap"])
     # Pre-compute per-symbol warmup cutoff dates (bar_idx >= WARMUP_BARS 的首日)
     # 用于替代 per-factor hash join，改为 pivot 后 pandas 广播掩码置 NaN
     cutoffs = (
@@ -81,12 +81,20 @@ def run_batch_eval(
     close_wide.index = pd.to_datetime(close_wide.index)
     rets = _forward_return_matrices(close_wide)
 
+    # 保存symbols_count(删除bars前)
+    symbols_count = bars["vt_symbol"].n_unique()
+
+    # 内存减负:bars(原始+bar_idx ~1G)派生完毕即释放——NAS 8G 盒防 OOM 周期性被杀
+    import gc
+    del bars
+    gc.collect()
+
     universe_label = universe if symbols is None else "custom"
     eval_store.init_db(eval_db)
     if run_id is None:
         run_id = eval_store.create_run(
             eval_db, label=label, universe=universe_label,
-            symbols_count=bars["vt_symbol"].n_unique(), factors_total=len(factor_names),
+            symbols_count=symbols_count, factors_total=len(factor_names),
             start=start, end=end, params={"limit": limit, "symbols": symbols[:20] if symbols else None},  # 截断防 4400 只全量塞 JSON
         )
         done_before = set()
@@ -104,7 +112,7 @@ def run_batch_eval(
                 "factors_done": len(done_before),
                 "errors": [],
                 "elapsed_sec": round(time.time() - t0, 1),
-                "symbols_count": int(bars["vt_symbol"].n_unique()),
+                "symbols_count": int(symbols_count),
             }
 
     errors: list[str] = []
@@ -130,7 +138,7 @@ def run_batch_eval(
         "factors_done": len(done_before) + done,
         "errors": errors,
         "elapsed_sec": round(time.time() - t0, 1),
-        "symbols_count": int(bars["vt_symbol"].n_unique()),
+        "symbols_count": int(symbols_count),
     }