fix jinja template

Former-commit-id: 0ebf2e2ee23918d28b0cbb20ba456732d6eedfbb

src/llamafactory/data/template.py

@@ -338,7 +338,11 @@ def _convert_slots_to_jinja(slots: "SLOTS", tokenizer: "PreTrainedTokenizer", pl
 
 
 def _get_jinja_template(template: "Template", tokenizer: "PreTrainedTokenizer") -> str:
-    jinja_template = _convert_slots_to_jinja(template.format_prefix.apply(), tokenizer)
+    jinja_template = ""
+
+    prefix = _convert_slots_to_jinja(template.format_prefix.apply(), tokenizer)
+    if prefix:
+        jinja_template += "{{ " + prefix + " }}"
 
     if template.default_system:
         jinja_template += "{% set system_message = '" + _jinja_escape(template.default_system) + "' %}"

tests/data/test_supervised.py

@@ -17,6 +17,7 @@ import random
 
 import pytest
 from datasets import load_dataset
+from transformers import AutoTokenizer
 
 from llamafactory.data import get_dataset
 from llamafactory.hparams import get_train_args
@@ -48,10 +49,11 @@ def test_supervised(num_samples: int):
     tokenizer = tokenizer_module["tokenizer"]
     tokenized_data = get_dataset(model_args, data_args, training_args, stage="sft", **tokenizer_module)
 
+    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
+
     original_data = load_dataset(TRAIN_ARGS["dataset"], split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)
     for index in indexes:
-        decoded_result = tokenizer.decode(tokenized_data["input_ids"][index])
         prompt = original_data[index]["instruction"]
         if original_data[index]["input"]:
             prompt += "\n" + original_data[index]["input"]
@@ -60,5 +62,6 @@ def test_supervised(num_samples: int):
             {"role": "user", "content": prompt},
             {"role": "assistant", "content": original_data[index]["output"]},
         ]
-        templated_result = tokenizer.apply_chat_template(messages, tokenize=False)
-        assert decoded_result == templated_result
+        templated_result = ref_tokenizer.apply_chat_template(messages, tokenize=False)
+        decoded_result = tokenizer.decode(tokenized_data["input_ids"][index])
+        assert templated_result == decoded_result

tests/data/test_template.py

@@ -0,0 +1,35 @@
+# Copyright 2024 the LlamaFactory team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+
+from transformers import AutoTokenizer
+
+from llamafactory.data import get_template_and_fix_tokenizer
+
+
+TINY_LLAMA = os.environ.get("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+
+
+def test_jinja_template():
+    tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
+    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
+    get_template_and_fix_tokenizer(tokenizer, name="llama3")
+    assert tokenizer.chat_template != ref_tokenizer.chat_template
+
+    messages = [
+        {"role": "user", "content": "hi!"},
+        {"role": "assistant", "content": "hello there"},
+    ]
+    assert tokenizer.apply_chat_template(messages) == ref_tokenizer.apply_chat_template(messages)