关于Amazon SageMaker GroundTruth私有工作组及AutoML的技术咨询

Hey there! Since you're new to AWS Ground Truth and working with private work teams for text classification, let's walk through your questions clearly:

• 1. Maximum number of annotators in a private work team
  AWS doesn't enforce a hard maximum limit on the number of members you can add to a private work team. The only practical constraints come from your AWS account's ability to create enough IAM or Cognito users (depending on how you set up your team). For large teams, you might want to consider task distribution efficiency, but technically, you can scale up as needed.

• 2. Pricing relation to private work team size
  No, Ground Truth pricing isn't tied to how many annotators are in your private team. You're charged based on the volume of completed annotations (e.g., each text entry labeled). If you use AutoML-assisted labeling, the pricing structure adjusts based on automated vs. human-reviewed labels, but the number of team members doesn't directly impact costs.

• 3. AutoML auto-labeling for new unannotated data
  AutoML won't automatically label new data just because you upload it to your S3 bucket. To reuse your existing labeled dataset for new unannotated data, follow these steps:

  1. First, train a custom model using your labeled dataset in Ground Truth (create a model training job targeting your text classification task).
  2. Once the model is trained and validated, create a new batch labeling job, selecting your trained model as the auto-labeling source.
  3. Point the job to your new unannotated data in S3, and Ground Truth will use the model to generate automated labels (you can also set up a human review step for low-confidence labels if needed).

• 4. 1000 objects requirement: total per category or overall?
  The documentation refers to 1000 labeled objects across all categories combined. That said, if your category distribution is extremely imbalanced (e.g., one category has only 50 labels while others have hundreds), AutoML model performance might suffer. For better results, aim for at least a few hundred labels per category, but the official minimum is 1000 total.

• 5. Maximum number of objects AutoML can label
  There's no hard upper limit on how many objects AutoML can label. The only constraints are your AWS account's resource quotas (like batch job concurrency) and the storage capacity of your S3 bucket. As long as you have a well-trained model (from your 1000+ manual labels), you can use it to process thousands or even millions of unannotated text entries. Just keep an eye on model confidence scores to ensure label quality stays high.

内容的提问来源于stack exchange，提问作者DGS