Skip to main content
Datasets

Datasets overview

Manage training datasets for fine-tuning models on QwenCloud.

Datasets are structured data files used to fine-tune models on QwenCloud. You can create, upload, and manage datasets from the Datasets console page.
Datasets are available only in Singapore and China North 2 (Beijing). They are not available in China (Hong Kong), Germany (Frankfurt), Japan (Tokyo), or US (Virginia), and their menu is hidden in the console in those regions. To use them, switch to Singapore or China North 2 (Beijing) with the region selector in the upper-right corner of the console. See Feature availability by region.
Datasets belong to the region where you create them. A dataset created in one region cannot be used in another region.

Data format

Datasets use JSONL format containing instruction-response pairs structured as a messages array. This format is used for SFT (Supervised Fine-Tuning) training.
FormatMax sizeRequirements
JSONL200 MBMust contain a messages array
You can upload up to 10 files per dataset, each no larger than 200 MB.

Data format example

Each line in the JSONL file must contain a messages array with role and content fields:
{
  "messages": [
    {"role": "system", "content": "You are a helpful assistant"},
    {"role": "user", "content": "Who painted the human body during the Renaissance?"},
    {"role": "assistant", "content": "<think>Optional model thinking</think>The Renaissance was a period of rebirth in art, culture, and scholarship. Many artists depicted the human form during this era."}
  ]
}
The <think> tag in the assistant response is optional. Include it if you want the fine-tuned model to produce chain-of-thought reasoning before its final answer.

Dataset workflow

When you create a dataset, it is saved as a Draft. You must publish the dataset before it can be used in a fine-tuning job. See Manage datasets for details.

Next steps