iTAG overview-Platform For AI(PAI)-阿里云帮助中心

iTAG is PAI's data labeling platform. It provides a range of templates for labeling image, text, video, audio, and multimodal data.

Important

Starting August 1, 2026, iTAG switches to allowlist-only access and enters maintenance mode with no further feature updates. New PAI customers, and customers who have not created labeling tasks or datasets in iTAG, cannot use iTAG. Customers who have already created labeling tasks or datasets can continue using the existing features. If you need AI-powered automatic dataset labeling, use the Multimodal Data Management feature.

Supported labeling tasks

iTAG provides the following preset labeling templates:

Image: image classification, object detection, image OCR, table recognition, and Image semantic segmentation.
Text: text classification, named entity recognition, and entity relation recognition.
Video: video classification, Video point annotation, and Video OCR.
Audio: audio classification, audio segmentation, and audio recognition.
Large model: Visual question answering, Multimodal RLHF annotation, Image-to-text, Image and text explanation, Conversation rewriting, Conversation sorting, and Conversation grouping.

Templates such as text classification and image classification are available directly in the console. To learn how to use other templates, see Template management.

Workflow

Create a dataset

Upload your data to OSS. Then use the dataset management module to import data from an OSS path and create a dataset. iTAG generates a .manifest index file (a JSONL file that contains data paths and metadata) for subsequent labeling tasks.

Important
Currently, iTAG supports only data stored in OSS. To ensure reliable access, the OSS bucket must be in the same region as PAI.
Create an annotation task
After you create a dataset, create and distribute labeling tasks in iTAG using a built-in or custom template. Task distribution involves three stages: labeling, review, and acceptance. Labeling is required, while review and acceptance are optional. The stages work as follows:
- Labeling: Annotators claim task packages on the labeling task page, complete the labeling work, and submit the results.
- Review: Annotators claim completed task packages on the review task page and review, modify, or reject the results.
- Acceptance: The requester claims task packages on the acceptance task page, performs a final review, and can approve, modify, or reject the results.
Process annotation tasks

Label, review, or accept task packages according to the task workflow to produce labeled data.
Export labeling results

Export the labeling results to a specified OSS directory for model training. The supported export formats are .csv, .jsonl, .manifest, and .coco.

Billing

iTAG platform (free): The iTAG platform is free of charge when your own team performs manual labeling.
Intelligent labeling service (free): The intelligent labeling service for certain large model labeling templates (such as image-to-text and image and text explanation) is currently free of charge. Any future charges will be announced separately.
Object Storage Service (charged): iTAG runs on Alibaba Cloud Object Storage Service (OSS). Fees incurred during use, such as OSS storage fees and data transfer fees, are billed separately according to OSS pricing.

Get help

Check the iTAG FAQ: If you encounter issues such as data loading errors, missing permissions, or OSS CORS configuration errors, see PAI-iTAG FAQ.
Ask the Alibaba Cloud AI assistant: The AI Assistant in the lower-right corner of the website answers your questions about using cloud products in real time. It provides technical support, product solutions, and documentation guidance.

Click the AI Assistant floating button in the lower-right corner of the page to open the chat panel and enter your question.