Create annotation jobs using preset templates for image, text, video, and audio data, or build custom templates for specific scenarios.
Prerequisites
Before you begin, ensure that you have:
An active PAI workspace. See Activate PAI and create a default workspace or Create and manage workspaces
Data uploaded to OSS and a dataset created in PAI. See Create a dataset for data annotation
Administrator or annotation administrator permissions. Contact a workspace administrator to grant this permission if needed. See Manage workspace members
How it works
Creating a labeling job in iTAG follows this sequence:
Select a dataset — Choose the OSS-backed dataset to annotate.
Choose a template — Pick a preset template (image, text, video, or audio) or build a custom template.
Configure labels — Define tag names and set single or multiple tag selection per object.
Set up the workflow — Assign subtask packages, configure the check proportion, and assign labelers and reviewers.
Publish — Click Create to publish the job. Labelers can then claim subtask packages from the Task Center.
Create a labeling job
Log on to the PAI console.
In the left-side navigation pane, click Workspaces. On the Workspaces page, click the name of the workspace you want to manage.
In the left-side navigation pane, choose Data Preparation > iTAG.
On the iTAG page, go to the Jobs tab and click Create Task.
On the Create Labeling Job page, configure the basic settings:
Parameter Description Task Name 1–100 characters. Must start with a letter, digit, or Chinese character. Can include underscores ( _) or hyphens (-).Input Dataset Select a dataset from PAI dataset management. Task Description A brief description to distinguish this job from others. Select a template type and template.
General Template: Use a preset platform template. Select one of the following:
Category Template Description Image OCR Recognize text within a selected image area. Object Detection Detect and locate objects in images. Image Classification Classify images with preset tags. PDF Perform OCR and classification on PDF files. Moderation and Matting Moderate and perform matting on images. Table Recognition Pre-recognize core table elements. Edit results as needed. Text Entity Recognition Recognize named entities in text. Text Classification Classify text with preset tags. Supports single-tag and multi-tag classification. Entity relationship Identify relationships between text entities for knowledge graph scenarios. Video Video Classification Classify videos with preset tags. Supports single-tag and multi-tag classification. Audio Audio Classification Classify audio files with preset tags. Supports single-tag and multi-tag classification. Audio segmentation Segment audio content and add tags to each segment. Audio Recognition Recognize and transcribe text from audio content. For application scenarios and data formats, see Image, Text, Video, and Audio.
Custom Template: See Labeling templates.
Custom Template: Build a template for a specific scenario by combining Content Components and Topic Components as prompted. For input and output data formats, see Custom templates.
(Conditional) If you selected Image > OCR, configure OCR recognition results configuration. The default value is OCR Recognition Results, which performs OCR on text within selected image areas.
Configure label settings under Label Configuration:
Enter tag names one at a time, pressing Enter after each name. For example, to label cats in images, add tags such as "Cat", "American Shorthair", and "British Shorthair".
Set the tag selection mode:
Single Choice: Each labeler applies one tag per selected object.
Multiple Choice: Each labeler applies multiple tags per selected object. For example, a labeler can select a cat and apply both "Cat" and "American Shorthair".
Single Choice and Multiple Choice control the number of tags applied to a single selected object, not the total number of selections per data sample.
(Optional) Enable intelligent pre-labeling under Enable Intelligent Labeling. See Data pre-labeling: Intelligent labeling configurations.
Configure subtask package allocation under Assign Subtask Packages. iTAG groups annotation tasks into subtask packages. Labelers claim packages and process all tasks within them. Choose one of the following allocation rules:
Rule Description Fixed size Assigns a fixed number of tasks per package. The allowed range depends on dataset size: 1–200 for datasets up to 20,000 records, 5–200 for 20,000–100,000 records, 25–200 for 100,000–500,000 records, and 50–200 for 500,000–1,000,000 records. Press Import Fields Groups data by a selected dataset field. Records with the same field value go into the same package. Targeted Assignment Assigns packages directly to specific labelers or teams. (Conditional) If your Task Workflow includes a check step — such as Marking-Check or Marking-Inspection-Acceptance — set the sampling ratio under Check Proportion. The default is 100%.
Under User Configuration, assign labelers, checkers, accepters, or task administrators (individuals or groups) based on the selected Task Workflow. For role permissions in iTAG, see iTAG overview.
Click Create.
View and manage labeling jobs
After creating a job, view all jobs on the Task Center page in iTAG. Check status and use the Actions column to manage jobs.
| Area | Task | Description |
|---|---|---|
| ① | Process annotation tasks | Click Go to the iTAG Page to access the annotation interface. Claim and process annotation, check, and acceptance tasks. See Process labeling tasks. |
| ② | View job status | View the status of all labeling jobs on the Task Center page. |
| ③ | Manage subtask packages | For incomplete jobs, click Subtask Details to review subtask package completion. For incomplete packages, click Transfer to reassign to another labeler, or Release to make the package available for others to claim. |
| ④ | Export results | For completed jobs, click Export Labeling Result to export results. Click Obtain Data Record to check retrieval progress. See Export labeling result data. |
| ⑤ | More operations | Click  in the Actions column to publish or unpublish a job. |
What's next
Claim and process annotation tasks. See Process labeling tasks.