[data] qwen3 fixes (#8109)

README.md

@@ -250,7 +250,7 @@ Compared to ChatGLM's [P-Tuning](https://github.com/THUDM/ChatGLM2-6B/tree/main/
 
 </details>
 
-> [!NOTE]
+> [!TIP]
 > If you cannot use the latest feature, please pull the latest code and install LLaMA-Factory again.
 
 ## Supported Models

README_zh.md

@@ -237,7 +237,7 @@ https://github.com/user-attachments/assets/43b700c6-a178-41db-b1f8-8190a5d3fcfc
 
 </details>
 
-> [!NOTE]
+> [!TIP]
 > 如果您无法使用最新的功能，请尝试重新拉取代码并再次安装 LLaMA-Factory。
 
 ## 模型

data/README.md

@@ -50,7 +50,9 @@ Currently we support datasets in **alpaca** and **sharegpt** format.
 
 * [Example dataset](alpaca_en_demo.json)
 
-In supervised fine-tuning, the `instruction` column will be concatenated with the `input` column and used as the human prompt, then the human prompt would be `instruction\ninput`. The `output` column represents the model response.
+In supervised fine-tuning, the `instruction` column will be concatenated with the `input` column and used as the user prompt, then the user prompt would be `instruction\ninput`. The `output` column represents the model response.
+
+For reasoning models, if the dataset contains chain-of-thought (CoT), the CoT needs to be placed in the model responses, such as `<think>cot</think>output`.
 
 The `system` column will be used as the system prompt if specified.
 
@@ -59,13 +61,13 @@ The `history` column is a list consisting of string tuples representing prompt-r
 ```json
 [
   {
-    "instruction": "human instruction (required)",
+    "instruction": "user instruction (required)",
-    "input": "human input (optional)",
+    "input": "user input (optional)",
     "output": "model response (required)",
     "system": "system prompt (optional)",
     "history": [
-      ["human instruction in the first round (optional)", "model response in the first round (optional)"],
+      ["user instruction in the first round (optional)", "model response in the first round (optional)"],
-      ["human instruction in the second round (optional)", "model response in the second round (optional)"]
+      ["user instruction in the second round (optional)", "model response in the second round (optional)"]
     ]
   }
 ]
@@ -86,6 +88,9 @@ Regarding the above dataset, the *dataset description* in `dataset_info.json` sh
 }
 ```
 
+> [!TIP]  
+> If the model has reasoning capabilities but the dataset does not contain chain-of-thought (CoT), LLaMA-Factory will automatically add empty CoT to the data. When `enable_thinking` is `True`, the empty CoT will be added to the model responses and loss computation will be considered; otherwise, it will be added to the user prompts and loss computation will be ignored. Please keep the `enable_thinking` parameter consistent during training and inference.
+
 ### Pre-training Dataset
 
 - [Example dataset](c4_demo.jsonl)
@@ -119,8 +124,8 @@ It requires a better response in `chosen` column and a worse response in `reject
 ```json
 [
   {
-    "instruction": "human instruction (required)",
+    "instruction": "user instruction (required)",
-    "input": "human input (optional)",
+    "input": "user input (optional)",
     "chosen": "chosen answer (required)",
     "rejected": "rejected answer (required)"
   }
@@ -174,7 +179,7 @@ Note that the human and observation should appear in odd positions, while gpt an
     "conversations": [
       {
         "from": "human",
-        "value": "human instruction"
+        "value": "user instruction"
       },
       {
         "from": "function_call",
@@ -225,7 +230,7 @@ Preference datasets in sharegpt format also require a better message in `chosen`
     "conversations": [
       {
         "from": "human",
-        "value": "human instruction"
+        "value": "user instruction"
       },
       {
         "from": "gpt",
@@ -233,7 +238,7 @@ Preference datasets in sharegpt format also require a better message in `chosen`
       },
       {
         "from": "human",
-        "value": "human instruction"
+        "value": "user instruction"
       }
     ],
     "chosen": {
@@ -275,7 +280,7 @@ KTO datasets require a extra `kto_tag` column containing the boolean human feedb
     "conversations": [
       {
         "from": "human",
-        "value": "human instruction"
+        "value": "user instruction"
       },
       {
         "from": "gpt",
@@ -314,7 +319,7 @@ The number of images should be identical to the `<image>` tokens in the conversa
     "conversations": [
       {
         "from": "human",
-        "value": "<image>human instruction"
+        "value": "<image>user instruction"
       },
       {
         "from": "gpt",
@@ -355,7 +360,7 @@ The number of videos should be identical to the `<video>` tokens in the conversa
     "conversations": [
       {
         "from": "human",
-        "value": "<video>human instruction"
+        "value": "<video>user instruction"
       },
       {
         "from": "gpt",
@@ -396,7 +401,7 @@ The number of audios should be identical to the `<audio>` tokens in the conversa
     "conversations": [
       {
         "from": "human",
-        "value": "<audio>human instruction"
+        "value": "<audio>user instruction"
       },
       {
         "from": "gpt",
@@ -437,7 +442,7 @@ The openai format is simply a special case of the sharegpt format, where the fir
       },
       {
         "role": "user",
-        "content": "human instruction"
+        "content": "user instruction"
       },
       {
         "role": "assistant",

data/README_zh.md

@@ -49,7 +49,9 @@
 
 - [样例数据集](alpaca_zh_demo.json)
 
-在指令监督微调时，`instruction` 列对应的内容会与 `input` 列对应的内容拼接后作为人类指令，即人类指令为 `instruction\ninput`。而 `output` 列对应的内容为模型回答。
+在指令监督微调时，`instruction` 列对应的内容会与 `input` 列对应的内容拼接后作为提示词，即提示词为 `instruction\ninput`。而 `output` 列对应的内容为模型回答。
+
+对于推理类模型的微调，如果数据集包含思维链，则需要把思维链放在模型回答中，例如 `<think>cot</think>output`。
 
 如果指定，`system` 列对应的内容将被作为系统提示词。
 
@@ -58,8 +60,8 @@
 ```json
 [
   {
-    "instruction": "人类指令（必填）",
+    "instruction": "用户指令（必填）",
-    "input": "人类输入（选填）",
+    "input": "用户输入（选填）",
     "output": "模型回答（必填）",
     "system": "系统提示词（选填）",
     "history": [
@@ -85,6 +87,9 @@
 }
 ```
 
+> [!TIP]
+> 如果模型本身具备推理能力，而数据集不包含思维链，LLaMA-Factory 会自动为数据添加空思维链。当 `enable_thinking` 为 `True` 时，空思维链会添加到模型回答中并且计算损失，否则会添加到用户指令中并且不计算损失。请在训练和推理时保持 `enable_thinking` 参数一致。
+
 ### 预训练数据集
 
 - [样例数据集](c4_demo.jsonl)
@@ -118,8 +123,8 @@
 ```json
 [
   {
-    "instruction": "人类指令（必填）",
+    "instruction": "用户指令（必填）",
-    "input": "人类输入（选填）",
+    "input": "用户输入（选填）",
     "chosen": "优质回答（必填）",
     "rejected": "劣质回答（必填）"
   }
@@ -173,7 +178,7 @@ KTO 数据集需要提供额外的 `kto_tag` 列。详情请参阅 [sharegpt](#s
     "conversations": [
       {
         "from": "human",
-        "value": "人类指令"
+        "value": "用户指令"
       },
       {
         "from": "function_call",
@@ -224,7 +229,7 @@ Sharegpt 格式的偏好数据集同样需要在 `chosen` 列中提供更优的
     "conversations": [
       {
         "from": "human",
-        "value": "人类指令"
+        "value": "用户指令"
       },
       {
         "from": "gpt",
@@ -232,7 +237,7 @@ Sharegpt 格式的偏好数据集同样需要在 `chosen` 列中提供更优的
       },
       {
         "from": "human",
-        "value": "人类指令"
+        "value": "用户指令"
       }
     ],
     "chosen": {
@@ -274,7 +279,7 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
     "conversations": [
       {
         "from": "human",
-        "value": "人类指令"
+        "value": "用户指令"
       },
       {
         "from": "gpt",
@@ -313,7 +318,7 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
     "conversations": [
       {
         "from": "human",
-        "value": "<image>人类指令"
+        "value": "<image><image>用户指令"
       },
       {
         "from": "gpt",
@@ -321,6 +326,7 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
       }
     ],
     "images": [
+      "图像路径（必填）",
       "图像路径（必填）"
     ]
   }
@@ -354,7 +360,7 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
     "conversations": [
       {
         "from": "human",
-        "value": "<video>人类指令"
+        "value": "<video><video>用户指令"
       },
       {
         "from": "gpt",
@@ -362,6 +368,7 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
       }
     ],
     "videos": [
+      "视频路径（必填）",
       "视频路径（必填）"
     ]
   }
@@ -395,7 +402,7 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
     "conversations": [
       {
         "from": "human",
-        "value": "<audio>人类指令"
+        "value": "<audio><audio>用户指令"
       },
       {
         "from": "gpt",
@@ -403,6 +410,7 @@ KTO 数据集需要额外添加一个 `kto_tag` 列，包含 bool 类型的人
       }
     ],
     "audios": [
+      "音频路径（必填）",
       "音频路径（必填）"
     ]
   }
@@ -437,7 +445,7 @@ OpenAI 格式仅仅是 sharegpt 格式的一种特殊情况，其中第一条消
       },
       {
         "role": "user",
-        "content": "人类指令"
+        "content": "用户指令"
       },
       {
         "role": "assistant",

scripts/vllm_infer.py

@@ -49,6 +49,8 @@ def vllm_infer(
     max_new_tokens: int = 1024,
     repetition_penalty: float = 1.0,
     skip_special_tokens: bool = True,
+    default_system: Optional[str] = None,
+    enable_thinking: bool = True,
     seed: Optional[int] = None,
     pipeline_parallel_size: int = 1,
     image_max_pixels: int = 768 * 768,
@@ -74,6 +76,8 @@ def vllm_infer(
             cutoff_len=cutoff_len,
             max_samples=max_samples,
             preprocessing_num_workers=16,
+            default_system=default_system,
+            enable_thinking=enable_thinking,
             vllm_config=vllm_config,
             temperature=temperature,
             top_p=top_p,
@@ -127,14 +131,11 @@ def vllm_infer(
         lora_request = None
 
     # Store all results in these lists
-    all_prompts = []
+    all_prompts, all_preds, all_labels = [], [], []
-    all_preds = []
-    all_labels = []
 
     # Add batch process to avoid the issue of too many files opened
     for i in tqdm(range(0, len(train_dataset), batch_size), desc="Processing batched inference"):
         vllm_inputs, prompts, labels = [], [], []
-
         batch = train_dataset[i : min(i + batch_size, len(train_dataset))]
 
         for j in range(len(batch["input_ids"])):
@@ -176,15 +177,14 @@ def vllm_infer(
             )
 
         results = llm.generate(vllm_inputs, sampling_params, lora_request=lora_request)
-
         preds = [result.outputs[0].text for result in results]
 
         # Accumulate results
         all_prompts.extend(prompts)
         all_preds.extend(preds)
         all_labels.extend(labels)
-
         gc.collect()
+
     # Write all results at once outside the loop
     with open(save_name, "w", encoding="utf-8") as f:
         for text, pred, label in zip(all_prompts, all_preds, all_labels):

src/llamafactory/chat/hf_engine.py

@@ -104,10 +104,7 @@ class HuggingfaceEngine(BaseEngine):
             messages, mm_input_dict["images"], mm_input_dict["videos"], mm_input_dict["audios"], processor
         )
         paired_messages = messages + [{"role": "assistant", "content": ""}]
-        system = system or generating_args["default_system"]
+        prompt_ids, _ = template.encode_oneturn(tokenizer, paired_messages, system, tools)
-        enable_thinking = input_kwargs.pop("enable_thinking", None)
-        enable_thinking = enable_thinking if enable_thinking is not None else generating_args["enable_thinking"]
-        prompt_ids, _ = template.encode_oneturn(tokenizer, paired_messages, system, tools, enable_thinking)
         prompt_ids, _ = template.mm_plugin.process_token_ids(
             prompt_ids,
             None,

src/llamafactory/chat/sglang_engine.py

@@ -160,10 +160,7 @@ class SGLangEngine(BaseEngine):
             messages, images or [], videos or [], audios or [], self.processor
         )
         paired_messages = messages + [{"role": "assistant", "content": ""}]
-        system = system or self.generating_args["default_system"]
+        prompt_ids, _ = self.template.encode_oneturn(self.tokenizer, paired_messages, system, tools)
-        enable_thinking = input_kwargs.pop("enable_thinking", None)
-        enable_thinking = enable_thinking if enable_thinking is not None else self.generating_args["enable_thinking"]
-        prompt_ids, _ = self.template.encode_oneturn(self.tokenizer, paired_messages, system, tools, enable_thinking)
         prompt_length = len(prompt_ids)
 
         temperature: Optional[float] = input_kwargs.pop("temperature", None)

src/llamafactory/chat/vllm_engine.py

@@ -124,10 +124,7 @@ class VllmEngine(BaseEngine):
             messages, images or [], videos or [], audios or [], self.processor
         )
         paired_messages = messages + [{"role": "assistant", "content": ""}]
-        system = system or self.generating_args["default_system"]
+        prompt_ids, _ = self.template.encode_oneturn(self.tokenizer, paired_messages, system, tools)
-        enable_thinking = input_kwargs.pop("enable_thinking", None)
-        enable_thinking = enable_thinking if enable_thinking is not None else self.generating_args["enable_thinking"]
-        prompt_ids, _ = self.template.encode_oneturn(self.tokenizer, paired_messages, system, tools, enable_thinking)
         prompt_length = len(prompt_ids)
 
         temperature: Optional[float] = input_kwargs.pop("temperature", None)

src/llamafactory/data/template.py

@@ -52,6 +52,7 @@ class Template:
     efficient_eos: bool
     replace_eos: bool
     replace_jinja_template: bool
+    enable_thinking: bool
     mm_plugin: "BasePlugin"
 
     def encode_oneturn(
@@ -60,7 +61,6 @@ class Template:
         messages: list[dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        enable_thinking: bool = False,
     ) -> tuple[list[int], list[int]]:
         r"""Return a single pair of token ids representing prompt and response respectively."""
         encoded_messages = self._encode(tokenizer, messages, system, tools)
@@ -94,7 +94,7 @@ class Template:
 
         return list(stop_token_ids)
 
-    def add_thought(self, content: str) -> str:
+    def add_thought(self, content: str = "") -> str:
         r"""Add empty thought to assistant message."""
         return f"{self.thought_words[0]}\n\n{self.thought_words[1]}\n\n" + content
 
@@ -105,7 +105,7 @@ class Template:
 
     def get_thought_word_ids(self, tokenizer: "PreTrainedTokenizer") -> list[int]:
         r"""Get the token ids of thought words."""
-        return tokenizer.encode(f"{self.thought_words[0]}\n\n{self.thought_words[1]}\n\n", add_special_tokens=False)
+        return tokenizer.encode(self.add_thought(), add_special_tokens=False)
 
     def _convert_elements_to_ids(self, tokenizer: "PreTrainedTokenizer", elements: "SLOTS") -> list[int]:
         r"""Convert elements to token ids."""
@@ -406,26 +406,21 @@ class ReasoningTemplate(Template):
         messages: list[dict[str, str]],
         system: Optional[str] = None,
         tools: Optional[str] = None,
-        enable_thinking: bool = False,
     ) -> tuple[list[int], list[int]]:
         messages = deepcopy(messages)
-        for i in range(len(messages)):
+        for i in range(1, len(messages) - 2, 2):
-            if messages[i]["role"] == Role.ASSISTANT and (i != len(messages) - 1):
             messages[i]["content"] = self.remove_thought(messages[i]["content"])
 
-        encoded_messages = self._encode(tokenizer, messages, system, tools)
+        prompt_ids, response_ids = super().encode_oneturn(tokenizer, messages, system, tools)
-        prompt_ids = []
+        if (
-        for encoded_ids in encoded_messages[:-1]:
+            self.thought_words[0] not in messages[-1]["content"]
-            prompt_ids += encoded_ids
-
-        if not enable_thinking and (
-            messages[-1]["role"] == Role.ASSISTANT
-            and self.thought_words[0] not in messages[-1]["content"]
             and self.thought_words[1] not in messages[-1]["content"]
         ):
-            prompt_ids += self.get_thought_word_ids(tokenizer)
+            if not self.enable_thinking:
+                prompt_ids = prompt_ids + self.get_thought_word_ids(tokenizer)
+            else:
+                response_ids = self.get_thought_word_ids(tokenizer) + response_ids
 
-        response_ids = encoded_messages[-1]
         return prompt_ids, response_ids
 
     @override
@@ -436,15 +431,16 @@ class ReasoningTemplate(Template):
         system: Optional[str] = None,
         tools: Optional[str] = None,
     ) -> list[tuple[list[int], list[int]]]:
-        messages = deepcopy(messages)
         encoded_messages = self._encode(tokenizer, messages, system, tools)
-        for i in range(len(messages) - 1):
+        for i in range(0, len(messages), 2):
             if (
-                messages[i + 1]["role"] == Role.ASSISTANT
+                self.thought_words[0] not in messages[i + 1]["content"]
-                and self.thought_words[0] not in messages[i + 1]["content"]
                 and self.thought_words[1] not in messages[i + 1]["content"]
             ):
+                if not self.enable_thinking:
                     encoded_messages[i] += self.get_thought_word_ids(tokenizer)
+                else:
+                    encoded_messages[i + 1] = self.get_thought_word_ids(tokenizer) + encoded_messages[i + 1]
 
         return [(encoded_messages[i], encoded_messages[i + 1]) for i in range(0, len(encoded_messages), 2)]
 
@@ -467,6 +463,7 @@ def register_template(
     efficient_eos: bool = False,
     replace_eos: bool = False,
     replace_jinja_template: bool = False,
+    enable_thinking: bool = True,
     mm_plugin: "BasePlugin" = get_mm_plugin(name="base"),
     template_class: type["Template"] = Template,
 ) -> None:
@@ -513,6 +510,7 @@ def register_template(
         efficient_eos=efficient_eos,
         replace_eos=replace_eos,
         replace_jinja_template=replace_jinja_template,
+        enable_thinking=enable_thinking,
         mm_plugin=mm_plugin,
     )
 
@@ -549,6 +547,7 @@ def parse_template(tokenizer: "PreTrainedTokenizer") -> "Template":
     messages = [{"role": "user", "content": "{{content}}"}, {"role": "assistant", "content": "{{content}}"}]
     assistant_slot = tokenizer.apply_chat_template(messages, add_generation_prompt=False, tokenize=False)
     assistant_slot = assistant_slot[len(prefix) + len(user_slot) :]
+    template_class = ReasoningTemplate if "<think>" in assistant_slot else Template
     assistant_slot = assistant_slot.replace("<think>", "").replace("</think>", "").lstrip("\n")  # remove thought tags
 
     if len(user_slot) > len(user_slot_empty_system):
@@ -558,7 +557,7 @@ def parse_template(tokenizer: "PreTrainedTokenizer") -> "Template":
     else:  # if defaut_system is empty, user_slot_empty_system will be longer than user_slot
         default_system = ""
 
-    return Template(
+    return template_class(
         format_user=StringFormatter(slots=[user_slot]),
         format_assistant=StringFormatter(slots=[assistant_slot]),
         format_system=StringFormatter(slots=[system_slot]),
@@ -572,6 +571,7 @@ def parse_template(tokenizer: "PreTrainedTokenizer") -> "Template":
         efficient_eos=False,
         replace_eos=False,
         replace_jinja_template=False,
+        enable_thinking=True,
         mm_plugin=get_mm_plugin(name="base"),
     )
 
@@ -600,6 +600,11 @@ def get_template_and_fix_tokenizer(tokenizer: "PreTrainedTokenizer", data_args:
         template.format_function = FunctionFormatter(slots=default_slots, tool_format=data_args.tool_format)
         template.format_tools = ToolFormatter(tool_format=data_args.tool_format)
 
+    if data_args.default_system is not None:
+        logger.info_rank0(f"Using default system message: {data_args.default_system}.")
+        template.default_system = data_args.default_system
+
+    template.enable_thinking = data_args.enable_thinking
     template.fix_special_tokens(tokenizer)
     template.fix_jinja_template(tokenizer)
     return template

src/llamafactory/hparams/data_args.py

@@ -115,6 +115,14 @@ class DataArguments:
         default=None,
         metadata={"help": "Tool format to use for constructing function calling examples."},
     )
+    default_system: Optional[str] = field(
+        default=None,
+        metadata={"help": "Override the default system message in the template."},
+    )
+    enable_thinking: bool = field(
+        default=True,
+        metadata={"help": "Whether or not to enable thinking mode for reasoning models."},
+    )
     tokenized_path: Optional[str] = field(
         default=None,
         metadata={

src/llamafactory/hparams/generating_args.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from dataclasses import asdict, dataclass, field
-from typing import Any, Optional
+from typing import Any
 
 from transformers import GenerationConfig
 
@@ -62,18 +62,10 @@ class GeneratingArguments:
         default=1.0,
         metadata={"help": "Exponential penalty to the length that is used with beam-based generation."},
     )
-    default_system: Optional[str] = field(
-        default=None,
-        metadata={"help": "Default system message to use in chat completion."},
-    )
     skip_special_tokens: bool = field(
         default=True,
         metadata={"help": "Whether or not to remove special tokens in the decoding."},
     )
-    enable_thinking: bool = field(
-        default=True,
-        metadata={"help": "Whether or not to enable thinking mode for reasoning models."},
-    )
 
     def to_dict(self, obey_generation_config: bool = False) -> dict[str, Any]:
         args = asdict(self)

src/llamafactory/webui/chatter.py

@@ -15,6 +15,7 @@
 import json
 import os
 from collections.abc import Generator
+from contextlib import contextmanager
 from typing import TYPE_CHECKING, Any, Optional
 
 from transformers.utils import is_torch_npu_available
@@ -68,6 +69,14 @@ def _format_response(text: str, lang: str, escape_html: bool, thought_words: tup
     )
 
 
+@contextmanager
+def update_attr(obj: Any, name: str, value: Any):
+    old_value = getattr(obj, name, None)
+    setattr(obj, name, value)
+    yield
+    setattr(obj, name, old_value)
+
+
 class WebChatModel(ChatModel):
     def __init__(self, manager: "Manager", demo_mode: bool = False, lazy_init: bool = True) -> None:
         self.manager = manager
@@ -198,6 +207,7 @@ class WebChatModel(ChatModel):
         Inputs: infer.chatbot, infer.messages, infer.system, infer.tools, infer.image, infer.video, ...
         Output: infer.chatbot, infer.messages
         """
+        with update_attr(self.engine.template, "enable_thinking", enable_thinking):
             chatbot.append({"role": "assistant", "content": ""})
             response = ""
             for new_text in self.stream_chat(
@@ -211,7 +221,6 @@ class WebChatModel(ChatModel):
                 top_p=top_p,
                 temperature=temperature,
                 skip_special_tokens=skip_special_tokens,
-            enable_thinking=enable_thinking,
             ):
                 response += new_text
                 if tools:

tests/data/test_template.py

@@ -126,29 +126,50 @@ def test_encode_multiturn(use_fast: bool):
 
 
 @pytest.mark.parametrize("use_fast", [True, False])
-def test_reasoning_encode_oneturn(use_fast: bool):
+@pytest.mark.parametrize("cot_messages", [True, False])
+@pytest.mark.parametrize("enable_thinking", [True, False])
+def test_reasoning_encode_oneturn(use_fast: bool, cot_messages: bool, enable_thinking: bool):
+    messages = MESSAGES_WITH_THOUGHT if cot_messages else MESSAGES
     tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B", use_fast=use_fast)
-    template = get_template_and_fix_tokenizer(tokenizer, DataArguments(template="qwen3"))
+    data_args = DataArguments(template="qwen3", enable_thinking=enable_thinking)
-    prompt_ids, answer_ids = template.encode_oneturn(tokenizer, MESSAGES)
+    template = get_template_and_fix_tokenizer(tokenizer, data_args)
+    prompt_ids, answer_ids = template.encode_oneturn(tokenizer, messages)
     prompt_str = (
-        "<|im_start|>user\nHow are you<|im_end|>\n"
+        f"<|im_start|>user\n{messages[0]['content']}<|im_end|>\n<|im_start|>assistant\n"
-        "<|im_start|>assistant\nI am fine!<|im_end|>\n"
+        f"{MESSAGES[1]['content']}<|im_end|>\n"
-        "<|im_start|>user\n你好<|im_end|>\n"
+        f"<|im_start|>user\n{messages[2]['content']}<|im_end|>\n<|im_start|>assistant\n"
-        "<|im_start|>assistant\n<think>\n\n</think>\n\n"
     )
-    answer_str = "很高兴认识你！<|im_end|>\n"
+    answer_str = f"{messages[3]['content']}<|im_end|>\n"
+    if not cot_messages:
+        if enable_thinking:
+            answer_str = "<think>\n\n</think>\n\n" + answer_str
+        else:
+            prompt_str = prompt_str + "<think>\n\n</think>\n\n"
+
     _check_tokenization(tokenizer, (prompt_ids, answer_ids), (prompt_str, answer_str))
 
 
 @pytest.mark.parametrize("use_fast", [True, False])
-def test_reasoning_encode_multiturn(use_fast: bool):
+@pytest.mark.parametrize("cot_messages", [True, False])
+@pytest.mark.parametrize("enable_thinking", [True, False])
+def test_reasoning_encode_multiturn(use_fast: bool, cot_messages: bool, enable_thinking: bool):
+    messages = MESSAGES_WITH_THOUGHT if cot_messages else MESSAGES
     tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B", use_fast=use_fast)
-    template = get_template_and_fix_tokenizer(tokenizer, DataArguments(template="qwen3"))
+    data_args = DataArguments(template="qwen3", enable_thinking=enable_thinking)
-    encoded_pairs = template.encode_multiturn(tokenizer, MESSAGES)
+    template = get_template_and_fix_tokenizer(tokenizer, data_args)
-    prompt_str_1 = "<|im_start|>user\nHow are you<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
+    encoded_pairs = template.encode_multiturn(tokenizer, messages)
-    answer_str_1 = "I am fine!<|im_end|>\n"
+    prompt_str_1 = f"<|im_start|>user\n{messages[0]['content']}<|im_end|>\n<|im_start|>assistant\n"
-    prompt_str_2 = "<|im_start|>user\n你好<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
+    answer_str_1 = f"{messages[1]['content']}<|im_end|>\n"
-    answer_str_2 = "很高兴认识你！<|im_end|>\n"
+    prompt_str_2 = f"<|im_start|>user\n{messages[2]['content']}<|im_end|>\n<|im_start|>assistant\n"
+    answer_str_2 = f"{messages[3]['content']}<|im_end|>\n"
+    if not cot_messages:
+        if enable_thinking:
+            answer_str_1 = "<think>\n\n</think>\n\n" + answer_str_1
+            answer_str_2 = "<think>\n\n</think>\n\n" + answer_str_2
+        else:
+            prompt_str_1 = prompt_str_1 + "<think>\n\n</think>\n\n"
+            prompt_str_2 = prompt_str_2 + "<think>\n\n</think>\n\n"
+
     _check_tokenization(
         tokenizer,
         (encoded_pairs[0][0], encoded_pairs[0][1], encoded_pairs[1][0], encoded_pairs[1][1]),
@@ -193,12 +214,12 @@ def test_get_stop_token_ids():
 @pytest.mark.parametrize("use_fast", [True, False])
 def test_gemma_template(use_fast: bool):
     prompt_str = (
-        "<bos><start_of_turn>user\nHow are you<end_of_turn>\n"
+        f"<bos><start_of_turn>user\n{MESSAGES[0]['content']}<end_of_turn>\n"
-        "<start_of_turn>model\nI am fine!<end_of_turn>\n"
+        f"<start_of_turn>model\n{MESSAGES[1]['content']}<end_of_turn>\n"
-        "<start_of_turn>user\n你好<end_of_turn>\n"
+        f"<start_of_turn>user\n{MESSAGES[2]['content']}<end_of_turn>\n"
         "<start_of_turn>model\n"
     )
-    answer_str = "很高兴认识你！<end_of_turn>\n"
+    answer_str = f"{MESSAGES[3]['content']}<end_of_turn>\n"
     _check_template("google/gemma-3-4b-it", "gemma", prompt_str, answer_str, use_fast)
 
 
@@ -206,12 +227,12 @@ def test_gemma_template(use_fast: bool):
 @pytest.mark.parametrize("use_fast", [True, False])
 def test_llama3_template(use_fast: bool):
     prompt_str = (
-        "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\nHow are you<|eot_id|>"
+        f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n{MESSAGES[0]['content']}<|eot_id|>"
-        "<|start_header_id|>assistant<|end_header_id|>\n\nI am fine!<|eot_id|>"
+        f"<|start_header_id|>assistant<|end_header_id|>\n\n{MESSAGES[1]['content']}<|eot_id|>"
-        "<|start_header_id|>user<|end_header_id|>\n\n你好<|eot_id|>"
+        f"<|start_header_id|>user<|end_header_id|>\n\n{MESSAGES[2]['content']}<|eot_id|>"
         "<|start_header_id|>assistant<|end_header_id|>\n\n"
     )
-    answer_str = "很高兴认识你！<|eot_id|>"
+    answer_str = f"{MESSAGES[3]['content']}<|eot_id|>"
     _check_template("meta-llama/Meta-Llama-3-8B-Instruct", "llama3", prompt_str, answer_str, use_fast)
 
 
@@ -220,12 +241,12 @@ def test_llama3_template(use_fast: bool):
 )
 def test_llama4_template(use_fast: bool):
     prompt_str = (
-        "<|begin_of_text|><|header_start|>user<|header_end|>\n\nHow are you<|eot|>"
+        f"<|begin_of_text|><|header_start|>user<|header_end|>\n\n{MESSAGES[0]['content']}<|eot|>"
-        "<|header_start|>assistant<|header_end|>\n\nI am fine!<|eot|>"
+        f"<|header_start|>assistant<|header_end|>\n\n{MESSAGES[1]['content']}<|eot|>"
-        "<|header_start|>user<|header_end|>\n\n你好<|eot|>"
+        f"<|header_start|>user<|header_end|>\n\n{MESSAGES[2]['content']}<|eot|>"
         "<|header_start|>assistant<|header_end|>\n\n"
     )
-    answer_str = "很高兴认识你！<|eot|>"
+    answer_str = f"{MESSAGES[3]['content']}<|eot|>"
     _check_template(TINY_LLAMA4, "llama4", prompt_str, answer_str, use_fast)
 
 
@@ -234,12 +255,12 @@ def test_llama4_template(use_fast: bool):
 )
 def test_phi4_template(use_fast: bool):
     prompt_str = (
-        "<|im_start|>user<|im_sep|>How are you<|im_end|>"
+        f"<|im_start|>user<|im_sep|>{MESSAGES[0]['content']}<|im_end|>"
-        "<|im_start|>assistant<|im_sep|>I am fine!<|im_end|>"
+        f"<|im_start|>assistant<|im_sep|>{MESSAGES[1]['content']}<|im_end|>"
-        "<|im_start|>user<|im_sep|>你好<|im_end|>"
+        f"<|im_start|>user<|im_sep|>{MESSAGES[2]['content']}<|im_end|>"
         "<|im_start|>assistant<|im_sep|>"
     )
-    answer_str = "很高兴认识你！<|im_end|>"
+    answer_str = f"{MESSAGES[3]['content']}<|im_end|>"
     _check_template("microsoft/phi-4", "phi4", prompt_str, answer_str, use_fast)
 
 
@@ -247,34 +268,30 @@ def test_phi4_template(use_fast: bool):
 def test_qwen2_5_template(use_fast: bool):
     prompt_str = (
         "<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n"
-        "<|im_start|>user\nHow are you<|im_end|>\n"
+        f"<|im_start|>user\n{MESSAGES[0]['content']}<|im_end|>\n"
-        "<|im_start|>assistant\nI am fine!<|im_end|>\n"
+        f"<|im_start|>assistant\n{MESSAGES[1]['content']}<|im_end|>\n"
-        "<|im_start|>user\n你好<|im_end|>\n"
+        f"<|im_start|>user\n{MESSAGES[2]['content']}<|im_end|>\n"
         "<|im_start|>assistant\n"
     )
-    answer_str = "很高兴认识你！<|im_end|>\n"
+    answer_str = f"{MESSAGES[3]['content']}<|im_end|>\n"
     _check_template("Qwen/Qwen2.5-7B-Instruct", "qwen", prompt_str, answer_str, use_fast)
 
 
 @pytest.mark.parametrize("use_fast", [True, False])
-def test_qwen3_template(use_fast: bool):
+@pytest.mark.parametrize("cot_messages", [True, False])
+def test_qwen3_template(use_fast: bool, cot_messages: bool):
+    messages = MESSAGES_WITH_THOUGHT if cot_messages else MESSAGES
     prompt_str = (
-        "<|im_start|>user\nHow are you<|im_end|>\n"
+        f"<|im_start|>user\n{messages[0]['content']}<|im_end|>\n"
-        "<|im_start|>assistant\nI am fine!<|im_end|>\n"
+        f"<|im_start|>assistant\n{MESSAGES[1]['content']}<|im_end|>\n"
-        "<|im_start|>user\n你好<|im_end|>\n"
+        f"<|im_start|>user\n{messages[2]['content']}<|im_end|>\n"
-        "<|im_start|>assistant\n<think>\n\n</think>\n\n"
-    )
-    answer_str = "很高兴认识你！<|im_end|>\n"
-    _check_template("Qwen/Qwen3-8B", "qwen3", prompt_str, answer_str, use_fast)
-
-    prompt_str = (
-        "<|im_start|>user\nHow are you<|im_end|>\n"
-        "<|im_start|>assistant\nI am fine!<|im_end|>\n"
-        "<|im_start|>user\n你好<|im_end|>\n"
         "<|im_start|>assistant\n"
     )
-    answer_str = "<think>\n模型思考内容\n</think>\n\n很高兴认识你！<|im_end|>\n"
+    answer_str = f"{messages[3]['content']}<|im_end|>\n"
-    _check_template("Qwen/Qwen3-8B", "qwen3", prompt_str, answer_str, use_fast, messages=MESSAGES_WITH_THOUGHT)
+    if not cot_messages:
+        answer_str = "<think>\n\n</think>\n\n" + answer_str
+
+    _check_template("Qwen/Qwen3-8B", "qwen3", prompt_str, answer_str, use_fast, messages=messages)
 
 
 def test_parse_llama3_template():
@@ -293,6 +310,7 @@ def test_parse_llama3_template():
 def test_parse_qwen_template():
     tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct", token=HF_TOKEN)
     template = parse_template(tokenizer)
+    assert template.__class__.__name__ == "Template"
     assert template.format_user.slots == ["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]
     assert template.format_assistant.slots == ["{{content}}<|im_end|>\n"]
     assert template.format_system.slots == ["<|im_start|>system\n{{content}}<|im_end|>\n"]
@@ -303,6 +321,7 @@ def test_parse_qwen_template():
 def test_parse_qwen3_template():
     tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B", token=HF_TOKEN)
     template = parse_template(tokenizer)
+    assert template.__class__.__name__ == "ReasoningTemplate"
     assert template.format_user.slots == ["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]
     assert template.format_assistant.slots == ["{{content}}<|im_end|>\n"]
     assert template.format_system.slots == ["<|im_start|>system\n{{content}}<|im_end|>\n"]