Creating a One-Click Training Job

Overview

For common tasks like text generation, one-click training is best if you do not want to handle scripts, hyperparameters, or resources. Just provide a dataset and choose a task. The platform picks the algorithms, sets up training, and allocates resources to create a ready-to-use model. This option offers "zero code and fast results," making it perfect for quick prototyping and fast service deployment.

Prerequisites

1. You have registered a Huawei Cloud account, completed real-name authentication, and been granted related permissions. For details, see 1. Preparations: Accounts and Permissions.
2. To use one-click model training, you need to configure the ModelArts FullAccess permission. Otherwise, you cannot access OBS. For details, see Configuring Common Operations Permissions for ModelArts.

Constraints

Only the new console supports one-click training. For details about how to switch between the old and new consoles, see New and Old Consoles.

Billing

Model training in ModelArts uses compute and storage resources, which are billed. Compute resources are billed for running training jobs. Storage resources are billed for storing data in OBS or SFS. For details, see Model Training Billing Items.

Procedure

1. Log in to the ModelArts console.
2. In the navigation pane, choose Getting Started. In the Popular Models area, select the model you want to train. Click Train on the model card. The page for creating a training job is displayed.
3. Quickly configure basic information.

   Table 1 Basic information

   Parameter	Description	Default
   Model	Create a model on which the training job depends.	Selected model.
   Job Name	Job name, which is mandatory. The system automatically generates a name, which you can then rename according to the following rules. The name contains 1 to 64 characters. Only letters, digits, hyphens (-), and underscores (_) are allowed.	A job name in the job-xxxx format is randomly generated.
   Training Method	Training method used to create a training job. Full fine-tuning and LoRA fine-tuning are available. Full fine-tuning: All model parameters are updated for new tasks. It works best but uses a lot of memory and compute. Use this method when you have plenty of data. LoRA fine-tuning: Freezes the original model's parameters and trains only the added low-rank adaptation matrix. This matrix adds only 0.1% to 1% extra parameters. It reduces memory and compute needs while keeping performance close to full fine-tuning. This method works well for systems with limited resources.	LoRA fine-tuning
   Training Framework	Training framework used to create a training job.	MindSpeed
   Model Output Path	Checkpoints and model assets generated by this training task will be saved to this path. 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/.	obs://bucketname/path/
   Resource Pool Type	Public resource pool: The public resource pool is available for all tenants and does not require user creation. Dedicated resource pool: Dedicated resource pools are created separately and used exclusively. For details, see Creating a Dedicated Resource Pool.	Public resource pool
   Resource Pool	This parameter appears only for dedicated resource pools. In the Resource Pool section, click Select Resource Pool and choose your desired dedicated resource pool or logical subpool from the menu on the right. Click OK. You can view the dedicated resource pool name, node pool specifications, number of available nodes/maximum number of nodes, number of available NPU/GPUs, available CPUs (vCPUs), available memory (GiB), and resource fragments. Hover over View in the Resource Fragment column to check fragment details and check whether the resource pool meets the training requirements. Once you choose a resource pool, its details appear. To choose a different one, click Reselect.	Select a resource pool that meets the training requirements.
   Asset Name	Name of the model asset to be published after training. Naming rules: The value can contain 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.	Original model name + Timestamp
   Instance Specifications	Resource instance specifications used to create a training job. Determines the hardware specifications for the training instances. For Dedicated resource pool, you must select a pool first. For Public resource pool, select a specification directly from the list. 8 x Ascend-snt9b2 | 192 vCPUs | 1536 GiB (modelarts.bm.npu.arm.8snt9b2) or higher is recommended.	8 x Ascend-snt9b2 | 192 vCPUs | 1536 GiB (modelarts.bm.npu.arm.8snt9b2)
   Instances	Number of resource instances used to create a training job. The value range displayed on the GUI is related to the instance specifications.	1
   Dataset	Dataset used for creating a training job. You can choose Preset Data or My Data. In this example, choose a ModelArts preset dataset.	First dataset option
   More Settings	Model parameters for creating a training job. For details, see Table 2.	The parameters are set by default.

   Table 2 More training parameters

   Parameter	Description	Default
   LR	Learning rate.	1.25e-6
   MIN_LR	Minimum learning rate.	1.25e-7
   RUN_TYPE	Training type, for example, lora or sft.	lora
   MODEL_NAME	Name of the model to train.	qwen3_8b
   SEQ_LEN	Maximum sequence length to process. The default value is 4096.	4096
   LR_WARMUP_RATIO	Ratio of LR_WARMUP_ITERS to TRAIN_ITERS. LR_WARMUP_ITERS=TRAIN_ITERS * LR_WARMUP_RATIO	0.01
   DATA_TYPE	Choose one of the following values based on the dataset: GeneralPretrainHandler: Use the pre-trained Alpaca dataset. AlpacaStyleInstructionHandler: Uses a LLaMA-Factory's Alpaca dataset. SharegptStyleInstructionHandler: Uses a LLaMA-Factory's ShareGPT dataset.	AlpacaStyleInstructionHandler
   EPOCH	Defines the number of training epochs, which should be modified based on requirements. An epoch is a process where all training samples are trained once.	4
   GBS	The total number of samples processed across all servers in one training step. It affects the duration of each iteration.	16
   LORA_R	In the LoRA matrix, the rank value is used to measure the complexity and information amount of the matrix. A larger value enhances the representation capability of the model, but increases the training duration. A smaller value reduces the number of parameters and the overfitting risk.	16

4. Confirm the configuration and price, and click OK. The model training job list page is displayed.

   A training job runs for a period of time. On the training job list page, view the basic information about the training job.

   • In the training job list, Status of a newly created training job is Pending.
   • When the status of a training job changes to Completed, the training job is finished, and the generated model is stored in the corresponding output path.
   • If the status of a training job changes to Failed or Abnormal, click the name of the fine-tuning job to go to the details page. View the model output, task details, events, logs, and resource usage to locate the fault.

After the model training is successful, you can directly use the model for subsequent operations such as model deployment, or create a model again on the training job creation page by referring to this training process as required.

FAQs

For details about common errors and solutions, see Table 3.

INFO: root: Train valid number is 0.

ValueError: label_map not match. {1:'apple', 2:'orange', 3:'banana', 4:'pear'} & {1:'apple', 2:'orange', 3:'banana'}