diff --git a/src/model/preprocessed_dataset.py b/src/model/preprocessed_dataset.py
index 13faf36..5f1eb51 100644
--- a/src/model/preprocessed_dataset.py
+++ b/src/model/preprocessed_dataset.py
@@ -93,7 +93,7 @@ class PreProcessedDataset(Dataset):
     所有数据以 int16 存储，读取时转为 torch.long (int64)。
     """
 
-    def __init__(self, data_dir: str, max_cache_shards: int = 2):
+    def __init__(self, data_dir: str, max_cache_shards: int = 1):
         self.data_dir = Path(data_dir)
 
         with open(self.data_dir / "metadata.json", "r", encoding="utf-8") as f:
@@ -144,12 +144,9 @@ class PreProcessedDataset(Dataset):
         self.pinyin_ids = np.load(self.data_dir / "pinyin_ids.npy", mmap_mode="r")
 
     def _load_shard(self, shard_idx: int) -> Dict[str, np.ndarray]:
-        """加载一个 .npz 分片到内存"""
+        """加载一个 .npz 分片到内存（保持原始 int16，不转换）"""
         shard_path = self.data_dir / f"shard_{shard_idx:06d}.npz"
-        data = dict(np.load(shard_path))
-        for key in data:
-            data[key] = data[key].astype(np.int64)
-        return data
+        return dict(np.load(shard_path))
 
     def __len__(self) -> int:
         return self.num_samples
@@ -173,22 +170,22 @@ class PreProcessedDataset(Dataset):
             shard_data = self._cache.get(shard_idx, self._load_shard)
             return {
                 "input_ids": torch.from_numpy(
-                    shard_data["input_ids"][local_idx].copy()
+                    shard_data["input_ids"][local_idx].astype(np.int64)
                 ),
                 "token_type_ids": torch.from_numpy(
-                    shard_data["token_type_ids"][local_idx].copy()
+                    shard_data["token_type_ids"][local_idx].astype(np.int64)
                 ),
                 "attention_mask": torch.from_numpy(
-                    shard_data["attention_mask"][local_idx].copy()
+                    shard_data["attention_mask"][local_idx].astype(np.int64)
                 ),
                 "labels": torch.tensor(
                     shard_data["labels"][local_idx], dtype=torch.long
                 ),
                 "history_slot_ids": torch.from_numpy(
-                    shard_data["history_slot_ids"][local_idx].copy()
+                    shard_data["history_slot_ids"][local_idx].astype(np.int64)
                 ),
                 "pinyin_ids": torch.from_numpy(
-                    shard_data["pinyin_ids"][local_idx].copy()
+                    shard_data["pinyin_ids"][local_idx].astype(np.int64)
                 ),
             }
         else:
diff --git a/src/model/trainer.py b/src/model/trainer.py
index 2fde42e..ffe65b7 100644
--- a/src/model/trainer.py
+++ b/src/model/trainer.py
@@ -1260,7 +1260,7 @@ def train(
     is_eval_preprocessed = is_preprocessed_data(eval_data_path)
 
     if is_train_preprocessed:
-        train_dataset = PreProcessedDataset(train_data_path)
+        train_dataset = PreProcessedDataset(train_data_path, max_cache_shards=1)
         total_steps = (len(train_dataset) // batch_size) * num_epochs
         train_dataloader = create_dataloader(
             dataset=train_dataset,
@@ -1292,7 +1292,7 @@ def train(
         config_table.add_row("数据", "训练数据类型", "流式数据")
 
     if is_eval_preprocessed:
-        eval_dataset = PreProcessedDataset(eval_data_path)
+        eval_dataset = PreProcessedDataset(eval_data_path, max_cache_shards=1)
         eval_dataloader = create_dataloader(
             dataset=eval_dataset,
             batch_size=batch_size,