diff --git a/PyTorch/built-in/nlp/Bert_Chinese_ID3433_for_PyTorch/requirements.txt b/PyTorch/built-in/nlp/Bert_Chinese_ID3433_for_PyTorch/requirements.txt
index e07523b532f34a513e3cc5336db4ae3607d3f236..28b43304e76f939f47fad944f07af80c5756cd23 100644
--- a/PyTorch/built-in/nlp/Bert_Chinese_ID3433_for_PyTorch/requirements.txt
+++ b/PyTorch/built-in/nlp/Bert_Chinese_ID3433_for_PyTorch/requirements.txt
@@ -1,6 +1,6 @@
 accelerate
-datasets >= 1.8.0
-tokenizers
+datasets == 1.18.3
+tokenizers == 0.11.1
 sentencepiece == 0.1.96
 protobuf == 3.14
 wikiextractor
diff --git a/PyTorch/built-in/rl/MAPPO_for_PyTorch/onpolicy/scripts/train/train_mpe.py b/PyTorch/built-in/rl/MAPPO_for_PyTorch/onpolicy/scripts/train/train_mpe.py
index a030eaeeb3ff5877dfafac4b12c1eeb132a87d87..9ef5cb82fd6c12a3956227c8c529e270d4f9a156 100644
--- a/PyTorch/built-in/rl/MAPPO_for_PyTorch/onpolicy/scripts/train/train_mpe.py
+++ b/PyTorch/built-in/rl/MAPPO_for_PyTorch/onpolicy/scripts/train/train_mpe.py
@@ -21,6 +21,9 @@ import numpy as np
 from pathlib import Path
 import torch
 import torch_npu
+if torch_npu.npu.utils.get_soc_version() == 251:
+    torch_npu.npu.config.allow_internal_format = True
+
 from torch_npu.contrib import transfer_to_npu
 from onpolicy.config import get_config
 from onpolicy.envs.mpe.MPE_env import MPEEnv
diff --git a/PyTorch/contrib/nlp/roberta_for_PyTorch/README.md b/PyTorch/contrib/nlp/roberta_for_PyTorch/README.md
index a33954a66ecb7a3fbc6db71044eafae5a6d96805..c8bb74298cb515a2dc213ec3469ee43a779c27b0 100644
--- a/PyTorch/contrib/nlp/roberta_for_PyTorch/README.md
+++ b/PyTorch/contrib/nlp/roberta_for_PyTorch/README.md
@@ -86,7 +86,7 @@ RoBERTa 在模型规模、算力和数据上，都比 BERT 有一定的提升。
 
 1. 获取数据集。
 
-   下载 `SST-2` 数据集，请参考 `examples/roberta/preprocess_GLUE_tasks.sh` 。
+   用户自行下载 `SST-2` 数据集，请参考 `examples/roberta/preprocess_GLUE_tasks.sh` 。
 
    `SST-2` 数据集目录结构参考如下所示。
 
@@ -114,7 +114,7 @@ RoBERTa 在模型规模、算力和数据上，都比 BERT 有一定的提升。
 
 2. 获取预训练模型
 
-    下载预训练模型 `RoBERTa.base` , 解压至源码包路径下：“./pre_train_model/RoBERTa.base/model.pt”。
+   用户自行下载预训练模型 `RoBERTa.base` , 解压至源码包路径下：“./pre_train_model/RoBERTa.base/model.pt”。
 
 
 # 开始训练