调整数据采样逻辑以提升模型训练效果

src/suinput/dataset.py

@@ -184,15 +184,15 @@ class PinyinInputDataset(IterableDataset):
         # 随机选择采样方式 (各1/3概率)
         choice = random.random()
 
-        if choice < 0.333:
+        if choice < 0.85:
             # 方式1: 靠近汉字的54个字符
             return context[-54:] if context_len >= 54 else context
-        elif choice < 0.667:
+        elif choice < 0.95:
             # 方式2: 随机位置取46个连续字符
             if context_len <= 46:
                 return context
             start = random.randint(0, context_len - 46)
-            return context[start : start + 46]
+            return context[start : start + 46] + context[-8:]
         else:
             # 方式3: 12+6×7组合
             if context_len < 12:

src/tmp_utils/gen_eval_dataset.py

@@ -31,7 +31,7 @@ if __name__ == "__main__":
         batch_size=1024,
         num_workers=1,
         worker_init_fn=worker_init_fn,
-        pin_memory=True if torch.cuda.is_available() else False,
+        # pin_memory=True if torch.cuda.is_available() else False,
         collate_fn=custom_collate_with_txt,
         prefetch_factor=8,
         persistent_workers=True,