Fix tokenizer: add <unk> to BPE vocab

tokenizer.json

@@ -36875,13 +36875,15 @@
   "model": {
     "type": "BPE",
     "dropout": null,
-    "unk_token": "[UNK]",
+    "unk_token": "<unk>",
     "continuing_subword_prefix": null,
     "end_of_word_suffix": null,
     "fuse_unk": false,
     "byte_fallback": false,
     "ignore_merges": false,
-    "vocab": {},
+    "vocab": {
+      "<unk>": 0
+    },
     "merges": []
   }
 }