Platform For AI:Fine-tune DeepSeek-R1 distill models

Supported models

Model Gallery supports LoRA supervised fine-tuning (SFT) for six distill models. The following table lists the recommended minimum computing resource configurations when you use the default hyperparameters and the provided dataset.

Quick start

1. Go to the Model Gallery page.

   1. Log on to the PAI console. In the left-side navigation pane, select your target Workspace.

   2. In the left-side navigation pane, choose QuickStart > Model Gallery.

      

2. On the Model Gallery page, search for and click the DeepSeek-R1-Distill-Qwen-7B model card to open the model details page. This page provides details about model training and deployment, including the required data format for SFT and model invocation methods.

   

3. Click Train in the upper-right corner. Configure the following key parameters:

   • Dataset configuration: This example uses the default dataset. You can also prepare a custom dataset according to the format requirements on the model details page and upload it to an Object Storage Service (OSS) bucket.

   • Model output path: Select an OSS path to store the fine-tuned model.

   • Computing Resources: For Source, select public resource. For Instance type, select ecs.gn7i-c16g1.4xlarge.

   • Hyperparameters: The following table describes the hyperparameters supported for LoRA supervised fine-tuning. You can adjust them as needed. For more information, see Fine-tuning guide for large language models.

     Hyperparameters

     Parameter	Type	Default (for 7B model)	Description
     learning_rate	float	5e-6	The learning rate, which controls the magnitude of model weight adjustments.
     num_train_epochs	int	6	The number of times the training dataset is iterated over.
     per_device_train_batch_size	int	2	The number of samples processed by each GPU in a single training iteration. A larger batch size can improve efficiency but also increases video memory requirements.
     gradient_accumulation_steps	int	2	The number of gradient accumulation steps.
     max_length	int	1024	The maximum token length of the input data that the model processes in a single training iteration.
     lora_rank	int	8	The LoRA dimension.
     lora_alpha	int	16	The LoRA scaling factor.
     lora_dropout	float	0	The dropout rate for LoRA training. It helps prevent overfitting by randomly dropping neurons during training.
     lorap_lr_ratio	float	16	The LoRA+ learning rate ratio (λ = ηB/ηA), where ηA and ηB are the learning rates for adapter matrices A and B, respectively. Compared to LoRA, LoRA+ uses different learning rates for the adapter matrices A and B to achieve better performance and faster fine-tuning without increasing computational requirements. Set lorap_lr_ratio to 0 to use standard LoRA instead of LoRA+.

4. Click Train. PAI automatically redirects you to the training job page where you can monitor the job status and view logs.

   

   When the training job succeeds, the system automatically registers the fine-tuned model in AI Asset Management - Models. You can then view or deploy the model. For details, see Register and manage models.

5. After the training is complete, click Deploy in the upper-right corner to deploy the fine-tuned model as an EAS service. The invocation method is the same as that for the original distill model. For more information, see the model details page or Deploy DeepSeek-V3 and DeepSeek-R1 models.

   

Billing

Model training in Model Gallery uses DLC. DLC is billed based on the duration of the training job. For more information, see Billing for DLC.

FAQ

Related topics

• Deploy DeepSeek-V3 and DeepSeek-R1 models

• Fine-tuning guide for large language models

• Model Gallery FAQ