Platform For AI:One-click fine-tuning of DeepSeek-R1 distill models

Supported models

PAI-Model Gallery supports LoRA supervised fine-tuning (SFT) training for the six distill models. The following table describes the recommended minimum configurations under the default parameters and datasets:

Train the model

1. Go to the Model Gallery page.

   1. Log on to the PAI console.

   2. In the upper-left corner, select a region based on your business requirements.

   3. In the left-side navigation pane, click Workspaces. On the Workspaces page, click the name of the workspace that you want to use.

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

2. On the Model Gallery page, click the DeepSeek-R1-Distill-Qwen-7B model card to go to the details page.

   This page provides detailed information on model deployment and training, such as the SFT data format and the invocation method.

   

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

   • Dataset Configuration: After you prepare the data, upload the data to the Object Storage Service (OSS) bucket.

   • Computing Resources: Choose suitable resources. The minimum configurations required under the default setup are listed in Supported models. If you need to adjust the hyperparameters, more video memory may be required.

   • Hyperparameters: The following table describes the hyperparameters supported by LoRA SFT. Adjust these based on your data and computing resources. For more information, see Guide to fine-tuning LLMs.

     Hyperparameter	Type	Default value (for 7B model as an example)	Description
     learning_rate	float	5e-6	The learning rate, which controls the magnitude of model weight adjustments.
     num_train_epochs	int	6	The number times the training dataset is reused.
     per_device_train_batch_size	int	2	The number of samples processed by each GPU in one training iteration. A higher value results in higher training efficiency and higher memory usage.
     gradient_accumulation_steps	int	2	The number of gradient accumulation steps.
     max_length	int	1024	The maximum token length of input data processed by the model in one training session.
     lora_rank	int	8	The LoRA dimension.
     lora_alpha	int	32	The LoRA weights.
     lora_dropout	float	0	The LoRA dropout rate. Randomly dropping neurons during the training process helps prevent overfitting.
     lorap_lr_ratio	float	16	The learning rate ratio in LoRA+ is defined as λ = ηB/ηA, where ηA and ηB​ are the learning rates for adapter matrices A and B, respectively. Compared to standard LoRA, LoRA+ allows for the use of different learning rates for critical parts of the process, leading to better performance and faster fine-tuning without increasing computational demands. When lorap_lr_ratio is set to 0, ithe standard LoRA is being used instead of LoRA+.

4. Click Train. You will be redirected to the model training page, and the training will begin. Here, you can view the status and logs of the training job.

   

   • If the training is successful, the model will be automatically registered in AI Asset Management - Models, where you can view or deploy it. For more information, see Register and manage models.

   • If the training fails, click next to Status to discover the cause or go to the Task log tab for more information. For common training errors and solutions, see Usage notes and FAQ about Model Gallery.

   • The Metric Curve section at the bottom of the training page displays the loss progression during training.

     

5. After successful training, click Deploy in the upper-right corner to deploy the trained model as an EAS service. The invocation method for the deployed model is the same as that of the original distill model. You can refer to the model detail page or One-click deployment of DeepSeek-V3 and DeepSeek-R1.

   

Billing

Training in Model Gallery uses the training capacities of Deep Learning Containers (DLC). DLC charges based on the duration of training jobs. After your training job ends, resource consumption will stop automatically and you don't need to stop it manually. Learn about Billing of DLC.

Usage notes

References

• One-click deployment of DeepSeek-V3 and DeepSeek-R1

• Guide to fine-tuning LLMs

• FAQ about Model Gallery