diff --git a/onnxruntime_extensions/_hf_cvt.py b/onnxruntime_extensions/_hf_cvt.py
index 4ba6bdc5f..5ac80c96c 100644
--- a/onnxruntime_extensions/_hf_cvt.py
+++ b/onnxruntime_extensions/_hf_cvt.py
@@ -54,8 +54,15 @@ def convert_json_vocab(hf_tokenizer):
         # get vocab object from json file
         vocab = tokenizer_json.get("model", {}).get("vocab", {})
         sorted_merges = tokenizer_json.get("model", {}).get("merges", [])
-        sorted_merges = [v_.replace("\n", "<0x0A>") for v_ in sorted_merges]
+        
         attrs = {"vocab": json.dumps(vocab, separators=(",", ":"))}
+
+        # merges data can be a list of string or list of list of string
+        if (all(isinstance(v_,(list,tuple)))  for v_ in sorted_merges) :
+            sorted_merges = [ " ".join(v if v != "\n" else "<0x0A>" for v in v_ ) for v_ in sorted_merges]
+        else : 
+            sorted_merges = [v_.replace("\n", "<0x0A>") for v_ in sorted_merges]
+
         attrs["merges"] = "\n".join(sorted_merges)
         if hf_tokenizer.added_tokens_encoder:
             token_map = [f"{_k}={_v}" for _k,
diff --git a/test/test_autotokenizer.py b/test/test_autotokenizer.py
index fc48edd23..ee5a77ed0 100644
--- a/test/test_autotokenizer.py
+++ b/test/test_autotokenizer.py
@@ -2,6 +2,7 @@
 # Licensed under the MIT License.
 import unittest
 
+import os
 import numpy as np
 from transformers import AutoTokenizer, GPT2Tokenizer
 from onnxruntime_extensions import OrtPyFunction, gen_processing_models, ort_inference, util
@@ -128,6 +129,19 @@ def print_prime(n):
         self.assertEqual(len(ids["input_ids"].shape), len(actual_ids.shape))
         np.testing.assert_array_equal(ids["input_ids"], actual_ids)
 
+    def test_microsoft_phi4(self):
+        script_dir = os.path.dirname(os.path.abspath(__file__))
+        tokenizer_dir = os.path.join(script_dir, 'data',"phi-4-mini-reasoning")
+
+        tokenizer = AutoTokenizer.from_pretrained(
+            tokenizer_dir,  torch_dtype="auto")
+        code = 'This is a sample Code'
+
+        ids = tokenizer(code, return_tensors="np")
+        ort_tok, _ = gen_processing_models(tokenizer, pre_kwargs={})
+        actual_ids, *_ = ort_inference(ort_tok, [code])
+        self.assertEqual(len(ids["input_ids"].shape), len(actual_ids.shape))
+        np.testing.assert_array_equal(ids["input_ids"], actual_ids)
 
 if __name__ == '__main__':
     unittest.main()