diff --git a/supervised_finetuning.py b/supervised_finetuning.py
index 12e234a..b88b5e0 100644
--- a/supervised_finetuning.py
+++ b/supervised_finetuning.py
@@ -523,16 +523,21 @@ def main():
     tokenizer = tokenizer_class.from_pretrained(tokenizer_name_or_path, **tokenizer_kwargs)
     prompt_template = get_conv_template(script_args.template_name)
     if tokenizer.eos_token_id is None:
-        tokenizer.eos_token = prompt_template.stop_str  # eos token is required for SFT
-        logger.info("Add eos token: {}".format(tokenizer.eos_token))
+        tokenizer.eos_token = prompt_template.stop_str  # eos token is required
+        tokenizer.add_special_tokens({"eos_token": tokenizer.eos_token})
+        logger.info(f"Add eos_token: {tokenizer.eos_token}, eos_token_id: {tokenizer.eos_token_id}")
+    if tokenizer.bos_token_id is None:
+        tokenizer.add_special_tokens({"bos_token": tokenizer.eos_token})
+        tokenizer.bos_token_id = tokenizer.eos_token_id
+        logger.info(f"Add bos_token: {tokenizer.bos_token}, bos_token_id: {tokenizer.bos_token_id}")
     if tokenizer.pad_token_id is None:
         if tokenizer.unk_token_id is not None:
             tokenizer.pad_token = tokenizer.unk_token
         else:
             tokenizer.pad_token = tokenizer.eos_token
-        logger.info("Add pad token: {}".format(tokenizer.pad_token))
-
+        logger.info(f"Add pad_token: {tokenizer.pad_token}, pad_token_id: {tokenizer.pad_token_id}")
     logger.debug(f"Tokenizer: {tokenizer}")
+
     IGNORE_INDEX = LabelSmoother.ignore_index if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
 
     # Get datasets