Overview of model training in Alibaba Cloud Model Studio: data management, model fine-tuning (SFT, DPO, CPT, RL, and OPD distillation training), model evaluation, model compression, and My Models, covering the complete workflow from data to model output.
Model training covers the complete workflow from training data preparation to model output management, corresponding to the Model Training menu in the Model Studio console.
- Data Management: Centrally manage training sets and evaluation sets, with import options such as local upload, OSS import, and log backflow, and clean and process data through data flows.
- model fine-tuning: Create training jobs in the console. It supports five training methods: SFT fine-tuning, DPO preference training, CPT continued pre-training, RL reinforcement learning, and OPD distillation training. You can also submit jobs via API/SDK. For fine-tuning of image, video, and speech synthesis models, and fine-tuning of decision models, see the corresponding documentation.
- Model Evaluation: Create custom and baseline evaluation jobs, view evaluation results, perform manual annotation, and check leaderboards to quantify model performance across evaluation dimensions.
- model compression: Compress models with quantification templates to produce lighter models.
- My Models: Centrally manage models produced by fine-tuning and models imported from OSS, with support for publishing and deployment.