The original link to the wikitext-2 dataset is invalid, so it has been changed to a source from Kaggle.both torch and mxnet.

Description of changes:

By submitting this pull request, I confirm that you can use, modify,
copy, and redistribute this contribution, under the terms of your
choice.

Author: whiteode

d2l/mxnet.py

@@ -2272,8 +2272,8 @@ def forward(self, tokens, segments, valid_lens=None, pred_positions=None):
         return encoded_X, mlm_Y_hat, nsp_Y_hat
 
 d2l.DATA_HUB['wikitext-2'] = (
-    'https://s3.amazonaws.com/research.metamind.io/wikitext/'
-    'wikitext-2-v1.zip', '3c914d17d80b1459be871a5039ac23e752a53cbe')
+    'https://www.kaggle.com/api/v1/datasets/download/bestwater/wikitext-2-v1'
+    '', 'ca5f319246c1e34d406780c0b6c5d1b0ec9b9a10')
 
 def _read_wiki(data_dir):
     """Defined in :numref:`sec_bert-dataset`"""
@@ -3104,7 +3104,11 @@ def download(url, folder='../data', sha1_hash=None):
         # For back compatability
         url, sha1_hash = DATA_HUB[url]
     os.makedirs(folder, exist_ok=True)
-    fname = os.path.join(folder, url.split('/')[-1])
+    file_name = url.split('/')[-1]
+    if (not "." in file_name) and file_name in ["wikitext-2-v1"]:
+        file_name += ".zip"
+    fname = os.path.join(folder, file_name)
+
     # Check if hit cache
     if os.path.exists(fname) and sha1_hash:
         sha1 = hashlib.sha1()