One-Stop Experience: Data Connection, Training, Inference, and Live Comparison

Scenario

In the field of AI development, developers often face a complex process from data preparation to model deployment, which is not only time-consuming but also prone to errors. Using traditional development tools and methods, developers often find it difficult to efficiently manage and optimize this process. Therefore, how to simplify this process and improve development efficiency has become an urgent issue to be addressed. ModelArts is a one-stop AI development platform for developers. It allows users to import their own datasets to the platform for tasks such as model training.

The dataset used in this use case is simple, and the training method is relatively basic. This case is for demonstration purposes only and aims to help users quickly understand how to use functions such as data connection, model training, model inference, and live comparison. This use case is not applicable to complex prediction scenarios. You can refer to this use case and adjust the dataset and training parameters as required to create a model service that meets your expectations.

Prerequisites

You have registered a Huawei Cloud account, completed real-name authentication, and been granted related permissions. For details, see 1. Prerequisites: Accounts and Permissions.
You have created an OBS bucket. For details, see Using OBS Console.

Constraints

Data Connection: The size of a data file or compressed package imported through a data connection cannot exceed 20 GB, and the total file size cannot exceed 20 GB.
Model Training: ModelArts does not support OBS buckets with bucket encryption enabled. Ensure this option is disabled when creating your OBS bucket.
Live Comparison: Only text generation LLMs are supported for live comparison. Models in other domains are not supported. A single comparison task supports a maximum of three models.

Billing

OBS storage: You are billed based on the storage capacity, storage class, and storage duration. For details, see OBS Billing Overview.
Deploying an inference service: You are charged based on the accumulated number of compute and storage resources. For details, see Billing Items for Inference Deployment (New Version).
Creating a training job: You are billed based on the total amount of compute and storage resources. For details, see Model Training Billing Items.

Workflow

Connect an existing dataset to ModelArts.
Train the model based on the Qwen3-8B model to optimize its performance.
Deploy the trained model as an executable inference service.
Compare the untrained and trained inference services in real time to view the model training effectiveness.
To avoid unnecessary billing, delete related resources as needed.

Step 1: Create a Data Connection

In data processing and model training scenarios, efficiently and accurately importing various datasets into ModelArts is essential to support subsequent data refining and model training. However, traditional data import methods have many limitations, such as the inability to customize task names and limited data format conversion capabilities. As a result, users face challenges in terms of inconvenient operations and low data processing efficiency when importing data. How to implement more flexible and efficient data import functions on a new platform has become an urgent issue for users. To address these challenges, ModelArts provides the data connection function, which supports the import of multiple common data types, meeting users' diverse requirements during the data preparation phase.

In the upper right corner, click Create Data Connection. On the displayed page, configure the parameters.

Some parameters are described as follows. For details about the parameters, see Creating a Data Connection.

**Table 1** Parameters required for creating a data connection
Parameter		Description	Example Value
Data Import	Dataset Type	Text, image, video, audio, and other types of datasets are supported. For details about the dataset format requirements, see Dataset Format Requirements.	Text Single-turn Q&A
	File format	The default value for single-turn Q&A is jsonl. You can click the format name to download the corresponding sample file. In this example, the sample file is used for training.	jsonl Alpaca Format Click Download Sample File.
	Connection Method	OBS and local upload are supported.	Local file
	Storage Location	The options are Object Storage Service – Bucket and Object Storage Service – Parallel File System. You can enter a storage address or click to select a storage location. The storage address must start with obs:// or / and end with a slash (/). It cannot contain double slashes (//) except in the prefix. For example, obs://bucketname/path/ or /bucketname/path/. In this example, select Object Storage Service – Bucket, enter the storage address, and click Add File to upload the downloaded sample file. The uploaded file will be saved to the OBS storage address.	Object Storage Service – Bucket obs://obstest/test/ single_qa-data-example_alpaca.jsonl
Generate Dataset	Dataset Name	Name of a custom dataset. The name must start with a letter and end with a letter or digit. It can contain 2 to 63 characters. Only letters, digits, hyphens (-), and underscores (_) are allowed.	dataset-qa1
Generate Dataset	Dataset Status	Only published datasets can be used by downstream tasks such as model training. If you select Publish Dataset, the generated dataset is in the Online state on the Asset Management > Data > My Data page and can be directly used by downstream model training jobs. If you do not select Publish Dataset, the generated dataset will be in the Offline state on the Asset Management > Data > My Data page and cannot be directly used by downstream model training jobs. You need to manually publish the dataset before using it.	Select Publish Dataset.

Click Create in the lower right corner.
On the Data Connections page, you can view the dataset task status. If the status is Success, the data connection task is successful.
Figure 1 Data connection
After the dataset is generated, choose Asset Management > Data in the navigation pane and view the generated dataset in the My Data tab. For more information, see My Data.
Figure 2 My data

Step 2: Train the Model Based on Qwen3-8B

In the context of large model training, fine-tuning refers to the process of performing secondary training on a pre-trained model using domain-specific datasets. This process updates the model weights, allowing it to adapt more effectively to specific task requirements. This phase enables the model to accurately perform tasks in specific scenarios such as copywriting, code generation, and professional Q&A.

In the navigation pane of the ModelArts console, choose Model Build > Training.

Click Create Training Job.

Some parameters are described as follows. For details about the parameters, see Creating a Training Job.

**Table 2** Parameters for creating a training job
Parameter		Description	Example Value
Training Mode		Fine-tuning job: applies to scenarios where you need to fine-tune existing pre-trained models, such as Qwen series. Low-threshold training: Use pre-configured high-quality model assets. There is no need to manage image building, environment dependencies, or code debugging, simply upload your training data and adjust key parameters. Custom job: applies to scenarios requiring full control over the training workflow, including the use of proprietary code or specialized images.	Fine-Tuning
Basic Information	Task Name	Name of a custom fine-tuning job. The name can contain 1 to 64 characters, including only letters, digits, hyphens (-), and underscores (_).	Qwen3-8B-training
Training Configuration	Select Model	Click a card to select a model from Built-in Models or My Models.	Built-in models Qwen3-8B V1.0.0
	Training Objective	ModelArts supports two fine-tuning types: full fine-tuning and LoRA fine-tuning. Different models support different fine-tuning types. Full fine-tuning: Updates all model parameters. Offers high accuracy but features slower convergence and longer training times. LoRA fine-tuning: Freezes the original model and injects trainable layers. Provides near-full parameter performance with fast convergence and shorter training times.	Full Fine-Tuning
	Model Output Path	Fine-tuned models can be stored in OBS and SFS Turbo. Currently, fine-tuned models can only be stored in OBS. Future updates will include support for SFS Turbo. You can choose your own OBS bucket or enter a path. The path must start with obs:// and end with a slash (/), like this: obs://bucketname/path/. For shared buckets from other users, you must enter the path.	OBS obs://obstest/test/output/
Resource Configuration	Resource Pool Type	Public resource pools are supported. The public resource pool is available for all tenants and does not require user creation.	Public resource pool
Data Configuration	Training Dataset	Click the card, choose My Data, select the dataset created in Step 1, and click OK.	My data dataset-qa1
publish model	Publish to Assets	After you select Auto publish final model , the trained model is automatically published to the Asset Management > My Models page on the console.	Select Auto publish final model.
	Publishing Method	New model: The published fine-tuned model is a new model and is displayed on the Asset Management > My Models page. New version of an existing model: The published fine-tuned model will be associated with an existing model on the Asset Management > Models > My Models page. Only the version number will change; you can view the updated version number within the model's details page.	New model
	Model Name	Sets the name for the newly generated model. It is used in Step 3. Enter 2 to 128 characters. Only letters, digits, hyphens (-), and underscores (_) are allowed. The name must start with a letter and end with a letter or digit.	Qwen3-8B-training

After setting the parameters, click Create Now to create a fine tuning job.
Generally, a fine-tuning job needs to run for a period of time. Once the fine-tuning job shows Completed, it has finished. The system saves the created model in model assets for later access.

Figure 3 Model training completed

Step 3: Deploy My Model

Model asset management is the core module of ModelArts, responsible for uniformly managing both Built-in Models and My Models. In this case, a model from My Models is deployed with few clicks and used for inference. If you need to use a custom image or configure other inference parameters, create a custom inference service. For details, see Performing Custom Service Deployment.

In the navigation pane of the ModelArts console, choose Asset Management > Models. Click the My Models tab. Click Deploy on the right of the model created in Step 2.

On the Create Service page, configure related information and click OK. In the billing notification dialog box, read related information and click OK.

Some parameters are described as follows. For details about the parameters, see Model Asset Management.

**Table 3** Parameters for creating a service
Parameter	Description	Example Value
Name	Name used to identify and manage the real-time service. Enter a name as prompted. Only letters, digits, hyphens (-), and underscores (_) are allowed. Enter 1 to 128 characters.	Qwen3-8B-training
Resource Pool Type	Public resource pools and dedicated resource pool are supported.	Public resource pool
Inference Unit	Select the hardware resource configuration for the real-time service instances.	Use the recommended value.
Auto Stop	Auto-stop timer. Default: 1 hour; maximum: 24 hours. When auto stop is enabled, the system tracks how long the service runs. It will shut down the service if the runtime goes beyond the set limit. After the real-time service is deployed, you can choose Model Inference > Real-Time Inference on the console. Choose More > Configure Auto Stop in the Operation column and reconfigure the auto stop time.	Select this option. Default: 1 hour.

In the navigation pane, choose Model Inference > Real-Time Inference to view the deployed model services.
If the service status is Running, the service is running properly.

Step 4: Perform Live Comparison

The live comparison feature provides an intuitive evaluation platform, allowing you to perform side-by-side benchmarking of different models using identical inputs.

The live comparison feature requires at least two models for comparison. Since a trained model has already been deployed, you need to deploy an untrained model for live comparison of model effectiveness. In this example, the Qwen3-8B-training and Qwen3-8B - Untrained models are used for comparison to verify the model training effect.

Deploying a Model for Comparison

In the navigation pane of the ModelArts console, choose Asset Management > Models. On the Built-in Models tab page, click Deploy on the Qwen3-8B model card.

On the Create Service page, configure related information and click OK. In the billing notification dialog box, read related information and click OK.

Some parameters are described as follows. For details about the parameters, see Model Asset Management.

**Table 4** Parameters for creating a service
Parameter	Description	Example Value
Name	Name used to identify and manage the real-time service. Enter a name as prompted. Only letters, digits, hyphens (-), and underscores (_) are allowed. Enter 1 to 128 characters.	Qwen3-8B - Untrained
Resource Pool Type	Public resource pools and dedicated resource pool are supported.	Public resource pool
Inference Unit	Select the hardware resource configuration for the real-time service instances.	Use the recommended value.
Auto Stop	Auto-stop timer. Default: 1 hour; maximum: 24 hours. When auto stop is enabled, the system tracks how long the service runs. It will shut down the service if the runtime goes beyond the set limit. After the real-time service is deployed, you can choose Model Inference > Real-Time Inference on the console. Choose More > Configure Auto Stop and reconfigure the auto stop time.	Select this option. Default: 1 hour.

In the navigation pane, choose Model Inference > Real-Time Inference to view the status of the deployed model service. If the service status is Running, the deployment is complete.

Live Comparison

Only services of the text generation type can be compared.

Log in to the ModelArts console. In the navigation pane on the left, choose Model Evaluation > Live Comparison.
In the upper right corner of the Live Comparison page, click Service Comparison. In the Live Comparison | Select Service dialog box, select two services, one trained and one untrained, and click OK.
Enter your question in the search box. Click or press Enter on the keyboard to send the question. Press Shift+Enter to start a new line. Compare the effectiveness of the two services.
For more information about live comparison, see Live Comparison.