Skip to main content

Exporting Fine-Tuned Models

After completion of training, you can export your converged model and push it to the intended platform space as a brand-new repository. Below are the basic export path instructions for mainstream frameworks.

Scheme A: Exporting via LLaMA-Factory​

After completing a training cycle in the Web GUI, select the [Export] tab in the navigation list to access configurables:

  1. Select the Best Checkpoint Choose the most optimal target version to transform from the list of training snapshots.

    checkpoint

  2. Establish Export Path and Target ID

    • Export dir: Local temporary directory path for storing the result.
    • CSGHub ID: The name of the target repository carrying the new fruits of labor. The format must embody the namespace (Example: your-username/brand-new-model-name).

    checkpoint

  3. Execute Export Execute with the export interface button. The system then merges the underlying framework and finetuned layers and automatically pushes to the new repository.

Scheme B: Parameter Configuration via MS-Swift CLI​

If using MS-Swift inside the console sandbox environment, execute equivalent export parameter commands:

  1. Define Backbone Directory Path (--model) Example:

    output/v0-20250715-175923/checkpoint-93/

  2. Check Merge LoRA Option To ensure the final result can operate independently in broad environments apart from stacking, it's recommended to merge all applied layer parameters uniformly.

    merge-lora

  3. Provide Push Target Parameters Complete the arguments to initialize the remote address specifying a concrete repository name (hub_model_id). Example:

    --push_to_hub true --hub_model_id xzgan001/Qwen2.5-0.5B-Instruct-self3 --use_hf true

    Tip: MS-Swift configurations cannot overwrite existing data. Therefore, you must specify a new, unoccupied model repository marker for --hub_model_id within the platform.

    Ms-swift

  4. Run Execute operation until the system triggers an upload complete log, then navigate to your platform homepage to confirm its delivery.