< img height="1" width="1" style="display:none;" alt="" src="https://px.ads.linkedin.com/collect/?pid=3131724&fmt=gif" />
Last updated:

    Create a Fine-tuning Instance

    Access Entry

    On the base model details page you want to fine-tune, click the Fine-tune Instance button in the top right corner to navigate to the creation page.

    Tip

    If the "Fine-tune Instance" button is not shown on the model details page, the model does not currently support this feature. Contact the platform administrator for more information.

    Configuration Parameters

    On the fine-tuning instance creation page, fill in the following configuration, then click Create Instance:

    Parameter Description
    Instance Name Custom name; must not duplicate existing instances (e.g., qwen-medical-finetune)
    Model ID The model identifier on the platform; defaults to the currently selected base model
    Region/Resource Config Select GPU resources appropriate for the base model size (VRAM must accommodate the model and training gradients)
    Runtime Framework Select the fine-tuning framework: LLaMA-Factory or MS-Swift

    View Instance List

    After creation, use the top navigation to open Model Training & Evaluation → Fine-tuning Instances to monitor instance startup progress and manage running fine-tuning tasks in real time. You can also view them centrally in Resource Management.

    Using the Fine-tuning Framework

    Once the instance is running, click the instance name to enter the fine-tuning framework interface:

    • LLaMA-Factory: Provides the visual LlamaBoard interface — select dataset, configure LoRA parameters, start training, and view real-time loss curves and other training metrics.
    • MS-Swift: Provides both command-line and web interface options, supporting more model types and quantized fine-tuning options.

    Exporting the Fine-tuned Model

    After training completes, merge and export the fine-tuned weights as a new independent model repository for subsequent inference deployment or evaluation.