Model Fine-tuning Overview
What is Model Fine-tuning
The platform provides GPU-accelerated fine-tuning instance hosting services, supporting mainstream fine-tuning frameworks (LLaMA-Factory, MS-Swift). Users only need to select compute resources and a dataset to quickly perform customized training on large models — no complex training code required.
Supported Fine-tuning Frameworks
| Framework | Features |
|---|---|
| LLaMA-Factory | Provides a visual web training UI (LlamaBoard); supports multiple fine-tuning methods (LoRA, QLoRA, full-parameter fine-tuning, etc.); easy to get started |
| MS-Swift | From ModelScope community; supports a wide range of domestic and international mainstream models; provides rich fine-tuning algorithms and quantization options |
Core Features
- One-Click Launch: Click the fine-tuning button on the model details page to automatically allocate compute and mount the model environment.
- Full Lifecycle Management: Covers fine-tuning creation, training metric monitoring, and model export.
- Visual Framework UI: LLaMA-Factory includes a visual training panel — configure training parameters without the command line.
- Flexible Compute: Select different GPU resource specifications on demand.
- Model Export: After fine-tuning, export the trained weights as an independent model repository.
Fine-tuning Workflow
Select base model → Create fine-tuning instance → Configure dataset and training parameters → Start training → Monitor training metrics → Export fine-tuned model