Training Services¶

Overview¶

Training services support training models from scratch and training or fine-tuning models with your own datasets. An easy-to-use interface and high-performance GPU clusters help users complete training tasks efficiently.

Training¶

Features¶

With training services, you can:

Configure GPU resources designed for efficient training.
Set up training workflows through an easy-to-use console.
Monitor training in real time through the console and logs, tracking metrics such as loss and accuracy.

Access the Training Console¶

Sign in to the platform console and open Training.
Review the task list, which shows task names, GPU configurations, GPU memory, models, resource usage, and available actions.

Create a Training Task¶

Step 1: Click the + button next to a task to create a new training task.
Step 2: Complete the task parameters listed below.
Step 3: Review the configuration and click Confirm to deploy the task.

Parameter	Description
Name	A unique name for the training task.
Model	The model used for training. Choose public or private models, such as meta-llama-3-8b-instruct.
Workers	The number of worker processes.
Shared memory	Shared memory size, such as 16 GiB.
Resources	GPU / CPU resource type and instance specifications, such as 2 × Tesla P40 with 24 GiB of GPU memory.
TimeZone / Image	The task time zone and runtime image (public / private).

Tip

Select a GPU count and GPU memory capacity appropriate for your training requirements.

New training job

Manage Training Tasks¶

View task details: Click a task name to view status, resource usage, and performance metrics.
Delete tasks: Click Delete for the corresponding task in the list.

Warning

Deletion is permanent and cannot be undone. Delete a task only after confirming that it is no longer needed.

TensorBoard integration: For supported tasks, click TensorBoard to visualize training metrics and logs.
Resource usage monitoring: Review the Resource column to monitor GPU hours consumed by each task and manage costs and resource allocation.

Fine-tune¶

Features¶

Fine-tuning provides a straightforward way to customize pretrained models with user-specific data. Through transfer learning, it can substantially reduce training time and improve accuracy in specialized domains. Key capabilities include:

Custom dataset integration: Easily upload and integrate your own datasets for fine-tuning.
Parameter customization: Adjust the learning rate, number of epochs, and other fine-tuning parameters.
Performance evaluation: Use evaluation tools to assess whether the model meets your requirements.
Resource optimization: Use the platform's infrastructure to optimize resource allocation, reducing overall training time and compute costs.

New fine-tuning job

Usage Steps¶

Step	Instructions
1. Select a pretrained model	Browse the model repository and choose a model based on its architecture, original training task, and compatibility with your dataset.
2. Upload a dataset	Provide the dataset location through Data Path. Ensure the data is correctly formatted and contains the required metadata. The platform provides a sample file through download example.json.
3. Configure fine-tuning parameters	Select Method: Full fine-tuning or LoRA (low-rank adaptation). Configure Epoch, Learning Rate, and Validation Ratio.
4. Evaluate performance	After fine-tuning, use evaluation tools to assess metrics such as accuracy, precision, recall, and F1 to confirm that requirements are met.

Parameter	Description
Learning Rate	The step size used to move toward a minimum of the loss function in each iteration.
Epochs	The number of times the learning algorithm processes the entire training dataset.
Batch Size	The number of training samples used in one iteration.
Validation Ratio	The proportion of training data set aside for validation.

Advanced Configuration¶

Advanced users can:

Adjust hyperparameters to improve model performance.
Integrate custom training scripts for a more tailored fine-tuning workflow.