Labeling templates-Platform For AI(PAI)-阿里云帮助中心

iTAG template types

iTAG provides the following three types of templates:

Official template (Recommended)

Official templates are pre-configured, standard labeling templates provided by Alibaba Cloud PAI. These templates cover common labeling scenarios such as Image-Text Explanation, conversation ranking, Visual Question Answering (VQA), and video OCR. You can use official templates directly for your labeling work, eliminating the need to design templates from scratch.
Tenant template

A tenant template is a personalized labeling template created and saved on the iTAG platform by a single tenant (an enterprise or team that uses the same account). These templates are customized for the tenant's own business scenarios and can be reused or shared with other project members within the tenant. This allows you to quickly start similar data labeling tasks and improves both efficiency and consistency.
Custom template

Custom templates allow you to create entirely new labeling specifications to meet the unique needs of your business. You can freely design and configure labeling templates on the iTAG platform. The custom template feature is especially important for tasks with special labeling requirements, as it ensures a highly flexible and precise data labeling process.

Create a labeling template

This topic demonstrates how to create an official template. The procedure is similar for tenant and custom templates.

Go to iTAG.
1. Log on to the PAI console.
2. In the left-side navigation pane, click Workspaces. On the page that appears, click the name of the workspace that you want to manage.
3. In the left-side navigation pane, choose Data Preparation>iTAG.
Click Go to the iTAG Page. Then, in the left-side navigation pane, choose Management Center>Asset Management.
On the Template Management tab, click Create Template, and set Template Type to Official Template.
Select a template and click Edit to open the template creation page.
Configure the parameters for the labeling template. For more information about the key parameters of each template, see Official templates.
Click Save Template to complete the configuration. You can also perform the following operations as needed:
- Click Back to return to the template selection page.
- Click Intelligent Configuration to configure services such as pre-labeling and UDFs.
- Click Deprecate Configuration to allow or disallow discarding samples.
- Click Import Dataset to change the dataset to be imported.
- Click Export Template to export the template in JSON format.
- Click Save Template to save the template with its current configuration.
- Click Create Task to save the configured template and create a corresponding task.

Next steps

Create a labeling task

You can create a labeling task in one of the following two ways:

On the Create Template page, click Create Task in the lower-right corner. Enter a name for the template and click Save. Then, on the Create Task page, configure the task parameters and click OK.
Create a labeling task from the Management Center page, which you can access by choosing Management Center > Task Management. This method provides more parameter options for greater customization than creating a task from the template page. For more details, see Create a labeling task.

Process a labeling task

For detailed instructions, see Process a labeling task.

Official templates

The iTAG template marketplace provides multiple official templates, organized into two main categories: large model and machine learning.

The following table describes some of the most common official templates. For details about other templates, refer to the information provided in the console.

Template name	Type	Tag	Description
Video Tagging	machine learning (video)	Video frame bounding box, Video frame marking, Global questions	The video tagging template detects content in videos. It provides three components: video frame bounding box (for drawing boxes), video frame marking (for annotating time ranges), and global questions (for adding labels to the entire video).
Video OCR	machine learning (video)	video OCR	The video OCR template recognizes text inside a bounding box drawn on a video.
Image Semantic Segmentation	machine learning (image)	image semantic segmentation	The image semantic segmentation template segments image content and lets you configure questions to label the segmented regions.
Image-to-Text	large model (multimodal)	Image-to-Text	The Image-to-Text template supports rapid image segmentation using SAM and proprietary segmentation algorithms. It combines these capabilities with the text generation power of large language models (LLMs) for AIGC-related annotation tasks.
Image-Text Explanation	large model (multimodal)	Image-Text Explanation	The Image-Text Explanation template links image selections to text entities.
Dialogue Rewriting	large model (large language model (LLM))	conversation rewriting	The conversation rewriting template allows you to ask questions about a given topic. You can select the most appropriate bot response and rewrite it.
Multimodal RLHF Annotation - Auto Q&A	large model (multimodal)	Multimodal RLHF Annotation	This template enables multimodal RLHF annotation. Building on conversation rewriting, it allows you to connect a bot for automatic question answering. You can also input multimodal content, such as images, in manual question-answering mode.
Multimodal RLHF Annotation - Manual Q&A	large model (multimodal)	Multimodal RLHF Annotation
Dialogue Ranking	large model (large language model (LLM))	conversation ranking	The conversation ranking template allows you to classify and rank conversational content.
Visual Question Answering - Multi-turn	large model (multimodal)	VQA Conversation	The Visual Question Answering (VQA) template lets you ask and answer questions about an image's content, a process that can be integrated with a bot UDF. A single-turn conversation ignores context, whereas multi-turn question answering generates responses based on the preceding conversation.
Visual Question Answering - Single-turn	large model (multimodal)	VQA Conversation
Dialogue Grouping	large model (large language model (LLM))	Dialogue Grouping	The Dialogue Grouping template groups dialogues into categories and supports dialogue rewriting.
Intelligent Scoring	large model (large language model (LLM))	Intelligent Scoring	The Intelligent Scoring template rewrites Q&A content. It also lets you configure different questions for each labeled item.