Alibaba Cloud Model Studio:Fine-tune Qwen

SFT training set

SFT uses training data in Chat Markup Language (ChatML) format, which supports multi-turn conversations and multiple role settings.

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

# A line of training data (JSON format), the typical structure is as follows 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 details about the system, user, and assistant roles, see Overview of text generation models. Sample training sets: SFT-ChatML_format_example.jsonl and SFT-ChatML_format_example.xlsx. XLS and XLSX formats support only single-turn conversations.

For a single training entry, all assistant rows support the "loss_weight" parameter, which sets the relative importance during training. The valid values range from 0.0 to 1.0. Higher values indicate greater importance.

This parameter is in invitational preview. To use it, you can 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 for thinking model

The training data supports multi-turn conversations and multiple role settings, but only the final assistant output is trained.

The \n characters before and after the think tags must be retained.

# A line of training data (JSON format), the typical structure is as follows 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> tags
   ...
  {"role": "user", "content": "User input 2"}, 
  {"role": "assistant", "content": "<think>\nExpected thinking content 2\n</think>\n\nExpected output 2"} --Thinking content can only be included in the final assistant output. 
]}

For details about the system, user, and assistant roles, see Overview of text generation models. Sample training set: SFT-deep_thinking_content_example.jsonl.

You can configure the model to omit the <think> tag in training samples. If you use this output method, do not 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 final assistant row of a single training entry supports the "loss_weight" parameter, which sets the relative importance during training. The valid values range from 0.0 to 1.0. Higher values indicate greater importance.

This parameter is in invitational preview. To use it, you can contact your account manager.

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

SFT for image understanding (Qwen-VL)

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

For more information about the differences between the system, user, and assistant roles, see Overview of text generation models. Sample training data in ChatML format:

# A line of training data (JSON format), the typical structure is as follows when expanded:
{"messages":[
  {"role":"user",
    "content":[
      {"text":"User input 1"},
      {"image":"Image file name 1"}]},
  {"role":"assistant",
    "content":[
      {"text":"Expected model output 1"}]},
  {"role":"user",
    "content":[
      {"text":"User input 2"}]},
  {"role":"assistant",
    "content":[
      {"text":"Expected model output 2"}]},
  ...
  ...
  ...
 ]}

The following are the requirements for ZIP files:

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

2. The training text data file must be named data.jsonl and placed in the root directory of the ZIP file. Ensure that the data.jsonl file appears immediately when you open the ZIP file.

3. A single image cannot exceed 1024 pixels in width or height. The maximum size is 10 MB. Supported formats: .bmp, .jpeg /.jpg, .png, .tif /.tiff, and .webp.

4. Image file names cannot be duplicated, even if the files are stored in different folders.

5. ZIP file directory structure:

   Single-level directory (recommended)

   The image files and the data.jsonl file reside in the root directory of the ZIP file.

   Trainingdata_vl.zip
      |--- data.jsonl # Note: Do not wrap in an outer folder
      |--- image1.png
      |--- image2.jpg

   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 can declare only the image file name, not the file path. For example:

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

   3. Image 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
       |--- png_folder
           └── image2.png

HTTP

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

curl -X POST https://dashscope-intl.aliyuncs.com/compatible-mode/v1/files \
-H "Authorization: Bearer $DASHSCOPE_API_KEY" \
--form 'file=@"path/to/your/sample.jsonl"' \
--form 'purpose="fine-tune"'

Response:

{
    "id": "file-ft-e73cafa11cef43a0ab75fb8e",
    "object": "file",
    "bytes": 23149,
    "filename": "qwen-fine-tune-sample.jsonl",
    "purpose": "fine-tune",
    "status": "processed",
    "created_at": 1769138847
}

HTTP

curl 'https://dashscope-intl.aliyuncs.com/api/v1/fine-tunes/<job_id>' \
--header 'Authorization: Bearer '${DASHSCOPE_API_KEY} \
--header 'Content-Type: application/json'

{
    "request_id": "d100cddb-ac85-4c82-bd5c-9b5421c5e94d",
    "output":
    {
        "job_id": "ft-202511272033-8ae7",
        "job_name": "ft-202511272033-8ae7",
        "status": "RUNNING",
        "finetuned_output": "qwen3-14b-ft-202511272033-8ae7",
        "model": "qwen3-14b",
        "base_model": "qwen3-14b",
        "training_file_ids":
        [
            "9e9ffdfa-c3bf-436e-9613-6f053c66aa6e"
        ],
        "validation_file_ids":
        [],
        "hyper_parameters":
        {
            "n_epochs": 1,
            "batch_size": 16,
            "learning_rate": "1.6e-5",
            "split": 0.9,
            "warmup_ratio": 0.0,
            "eval_steps": 1,
            "save_strategy": "epoch",
            "save_total_limit": 10
        },
        "training_type": "sft",
        "create_time": "2025-11-27 20:33:15",
        "workspace_id": "llm-8v53etv3hwb8orx1",
        "user_identity": "1654290265984853",
        "modifier": "1654290265984853",
        "creator": "1654290265984853",
        "group": "llm",
        "max_output_cnt": 10
    }
}

HTTP

curl 'https://dashscope-intl.aliyuncs.com/api/v1/fine-tunes/<job_id>/logs?offset=0&line=1000' \
--header 'Authorization: Bearer '${DASHSCOPE_API_KEY} \
--header 'Content-Type: application/json' 

Use the offset and line parameters to retrieve a specific range of logs. The offset parameter specifies the starting position and line specifies the maximum number of log entries.

Sample response:

{
    "request_id":"1100d073-4673-47df-aed8-c35b3108e968",
    "output":{
        "total":57,
        "logs":[
            "{Log output 1}",
            "{Log output 2}",
            ...
            ...
            ...
        ]
    }
}