Fine-tune with the API or CLI

SFT training set

SFT ChatML (Chat Markup Language) format training data supports multi-turn conversations and various role settings.

The OpenAI name and weight parameters are not supported. All assistant outputs will be trained.

# A single line of training data (in JSON format) has the following typical structure when expanded:
{"messages": [
  {"role": "system", "content": "System input 1"},
  {"role": "user", "content": "User input 1"},
  {"role": "assistant", "content": "Expected model output 1"},
  {"role": "user", "content": "User input 2"},
  {"role": "assistant", "content": "Expected model output 2"}
  ...
]}

For information about the differences between system, user, and assistant, see Overview. Sample training datasets: SFT-ChatML_format_example.jsonl, SFT-ChatML_format_example.xlsx. XLS and XLSX formats support only single-turn conversations.

All assistant lines in a single training data entry support the "loss_weight" parameter, which sets the relative importance of that line during training. (Range: 0.0 to 1.0. A larger value indicates higher importance.)

This parameter is available for invitational preview. To use it, contact your account manager.

 {"role": "assistant", "content": "Expected model output 1", "loss_weight": 1.0},
 {"role": "assistant", "content": "Expected model output 2", "loss_weight": 0.5}

SFT thinking model

The training data supports multi-turn conversations and various role settings, but only the last assistant output can be trained. A single line of training data has the following structure when expanded:

The newline characters \n before and after the think tag must be preserved.

# A single line of training data (in JSON format) has the following typical structure when expanded:
{"messages": [
  {"role": "system", "content": "System input 1"},
  {"role": "user", "content": "User input 1"},
  {"role": "assistant", "content": "Model output 1"}, --Intermediate assistant outputs should not have <think style={{ verticalAlign: "top" }}> tags
   ...
  {"role": "user", "content": "User input 2"},
  {"role": "assistant", "content": "<think style={{ verticalAlign: "top" }}>\nExpected thinking content 2\n</think>\n\nExpected output 2"} --Thinking content can only be included in the last assistant output.
]}

For information about the differences between system, user, and assistant, see Overview. Sample training dataset: SFT-deep_thinking_content_example.jsonl.

You can also set the model not to output the <think style={{ verticalAlign: "top" }}> tag in the training samples. If you use this output method, it is not recommended to enable the thinking mode for calls after the model is trained.

{"role": "assistant", "content": "Expected model output 2"}  --Tells the model not to enable thinking

The last assistant line of a single training data entry supports the "loss_weight" parameter, which sets the relative importance of that entry during training. (Range: 0.0 to 1.0. A larger value indicates higher importance.)

This parameter is available for invitational preview. To use it, contact your account manager.

 {"role": "assistant", "content": "<think style={{ verticalAlign: "top" }}>\nExpected thinking content 2\n</think>\n\nExpected output 2", "loss_weight": 1.0}

SFT visual understanding (Qwen-VL)

The OpenAI name and weight parameters are not supported. All assistant outputs will be trained.

For information about the differences between system, user, and assistant, see Overview. Sample ChatML format training data:

To pass a system message, the corresponding content must use the array format [{"text":"..."}] and not the string format "content":"string".

# A single line of training data (in JSON format) has the following structure when expanded:
{"messages": [
  {"role": "system", "content": [{"text": "System input"}]},
  {"role": "user", "content": [{"text": "User input 1"}, {"image": "image_file_name1.jpg", "resized_width": 200, "resized_height": 200}]},
  {"role": "assistant", "content": [{"text": "Expected model output 1"}]},
  {"role": "user", "content": [{"text": "User input 2"}, {"video": "video_file_name1.mp4", "fps": 3.0, "resized_width": 200, "resized_height": 200, "video_start": 0.0, "video_end": 3.0}]},
  {"role": "assistant", "content": [{"text": "Expected model output 2"}]},
  {"role": "user", "content": [{"text": "User input 2"}, {"video": ["0.jpg", "1.jpg", "2.jpg", "3.jpg"], "sample_fps": 5.0, "resized_width": 200, "resized_height": 200}]},
  {"role": "assistant", "content": [{"text": "Expected model output 2"}]},
  ...
]}

Suggestions for training object localization:

• Qwen2.5-VL: The training coordinates are absolute values relative to the top-left corner of the scaled image, in pixels.
• Qwen3-VL: The training coordinates are relative, and the coordinate values are scaled to the range of [0, 999].

ZIP file requirements:

1. Format: ZIP. The maximum size is 2 GB. Folders and file names within the ZIP file must only contain ASCII characters, including letters (a-z, A-Z), numbers (0-9), underscores (_), and hyphens (-).

2. The training text file must be named data.jsonl and located in the root directory of the ZIP file. Ensure that thedata.jsonlfile is at the top level of the ZIP file and not in a subfolder.

3. The width and height of a single image must not exceed 1024 pixels, and the size must not exceed 10 MB. The supported formats are .bmp, .jpeg /.jpg, .png, .tif /.tiff, and .webp.

4. Image file names must be unique, even if they are in different folders.

5. ZIP file directory structure:

   Single-level directory (recommended)

   The image files and the data.jsonl file are all located in the root directory of the ZIP file.

   Trainingdata_vl.zip
      |--- data.jsonl #Note: Do not wrap in an outer folder
      |--- image1.png
      |--- video1.mp4
   

   Multi-level directory

   1. The data.jsonl file must be in the root directory of the ZIP file.

   2. In the data.jsonl file, you only need to specify the image or video filename, not the file path. For example:

      Correct example: image1.jpg. Incorrect example: jpg_folder/image1.jpg.

   3. Image/video file names must be globally unique within the ZIP file.

   Trainingdata_vl.zip
       |--- data.jsonl #Note: Do not wrap in an outer folder
       |--- jpg_folder
       |   └── image1.jpg
       |--- mp4_folder
           └── video.mp4
   

DPO dataset

DPO ChatML format training data. A single line of training data has the following structure when expanded:

For information about the differences between system, user, and assistant, see Overview. Sample training dataset: DPO_ChatML_format_example.jsonl.

# A single line of training data (in JSON format) has the following typical structure when expanded:
{"messages": [
  {"role": "system", "content": "System input"},
  {"role": "user", "content": "User input 1"},
  {"role": "assistant", "content": "Model output 1"},
  {"role": "user", "content": "User input 2"},
  {"role": "assistant", "content": "Model output 2"},
  {"role": "user", "content": "User input 3"}
 ],
 "chosen":
   {"role": "assistant", "content": "Agreed-upon expected model output 3"},
 "rejected":
   {"role": "assistant", "content": "Rejected expected model output 3"}}

The model takes all content within messages as input. DPO is used to train the model on positive and negative feedback for User input 3.

For deep thinking content, you need to wrap it with <think style={{ verticalAlign: "top" }}> tags:

{"role": "assistant", "content": "<think style={{ verticalAlign: "top" }}>Expected model thinking content</think>Expected model output"}

The "chosen" module of a single training data entry supports the "loss_weight" parameter, which sets the relative importance of that training data in the training process. (Range: 0.0 to 1.0. A larger value indicates higher importance.)

This parameter is available for invitational preview. To use it, contact your account manager.

 "chosen":
   {"role": "assistant", "content": "Agreed-upon expected model output 3", "loss_weight": 1.0},

CPT training set

CPT plain text format training data. A single line of training data has the following structure when expanded:

{"text":"Text content"}

Sample training dataset: CPT-text_generation_training_set_example.jsonl

DashScope API

For Windows CMD, replace ${DASHSCOPE_API_KEY} with %DASHSCOPE_API_KEY%. For PowerShell, replace it with $env:DASHSCOPE_API_KEY.

curl --request POST \
'https://dashscope.aliyuncs.com/api/v1/files' \
--header 'Authorization: Bearer '${DASHSCOPE_API_KEY} \
--form 'files=@"/path/to/your/file.jsonl"' \
--form 'purpose="fine-tune"' \
--form 'descriptions="a sample fine-tune data file for qwen"'

Sample response:

{
  "request_id":"xx",
  "data":{
    "uploaded_files":[{
      "file_id":"976bd01a-f30b-4414-86fd-50c54486e3ef",
      "name":"qwen-fine-tune-sample.jsonl"}],
  　"failed_uploads":[]}
 }