Create a model evaluation job that uses human workers

Important

Custom IAM policies that allow Amazon SageMaker Studio or Amazon SageMaker Studio Classic to create Amazon SageMaker resources must also grant permissions to add tags to those resources. The permission to add tags to resources is required because Studio and Studio Classic automatically tag any resources they create. If an IAM policy allows Studio and Studio Classic to create resources but does not allow tagging, "AccessDenied" errors can occur when trying to create resources. For more information, see Provide permissions for tagging SageMaker resources.

AWS managed policies for Amazon SageMaker that give permissions to create SageMaker resources already include permissions to add tags while creating those resources.

To create a model evaluation job that uses human workers you must set up your environment to have the correct permissions. Then, you can use the model evaluation job wizard in Studio to select the models you want to use, and then define the parameters and the workforce you want to use in the model evaluation job.

When the job is complete you can, view a report to understand how your workforce evaluated the models you selected. The results are also saved in Amazon S3 as a jsonlines output file.

In model evaluation job that uses human workers, you have the ability to bring inference data from models hosted outside of SageMaker and models hosted outside of AWS. To learn more, see Using your own inference data in model evaluation jobs that use human workers.

When your jobs are completed the results are saved in the Amazon S3 bucket specified when the job was created. To learn how to interpret your results, see Understand the results of your model evaluation job.

Prerequisites

To run a model evaluation in the Amazon SageMaker Studio UI, your AWS Identity and Access Management (IAM) role and any input datasets must have the correct permissions. If you do not have a SageMaker Domain or IAM role, follow the steps in Guide to getting set up with Amazon SageMaker.

Setting up your permissions

The following section shows you how to create a Amazon S3 bucket and how to specify the correct Cross-origin resource sharing (CORS) permissions.

To create a Amazon S3 bucket and specify the CORS permissions

Open the Amazon SageMaker console at https://console.aws.amazon.com/sagemaker/.
In the navigation pane, enter S3 into the search bar at the top of the page.
Choose S3 under Services.
Choose Buckets from the navigation pane.
In the General purpose buckets section, under Name, choose the name of the S3 bucket that you want to use to store your model input and output in the console. If you do not have an S3 bucket, do the following.
1. Select Create bucket to open a new Create bucket page.
2. In the General configuration section, under AWS Region, select the AWS region where your foundation model is located.
3. Name your S3 bucket in the input box under Bucket name.
4. Accept all of the default choices.
5. Select Create bucket.
6. In the General purpose buckets section, under Name, select the name of the S3 bucket that you created.
Choose the Permissions tab.
Scroll to the Cross-origin resource sharing (CORS) section at the bottom of the window. Choose Edit.

The following is the minimum required CORS policy that you must add to your Amazon S3 bucket. Copy and paste the following into the input box.


[
{
    "AllowedHeaders": ["*"],
    "AllowedMethods": [
        "GET",
        "HEAD",
        "PUT"
    ],
    "AllowedOrigins": [
        "*"
    ],
    "ExposeHeaders": [
      "Access-Control-Allow-Origin"
    ],
    "MaxAgeSeconds": 3000
}
]

Choose Save changes.

To add permissions to your IAM policy

You may want to consider the level of permissions to attach to your IAM role.

You can create a custom IAM policy that allows the minimum required permissions tailored to this service.
You can attach the existing AmazonSageMakerFullAccess and AmazonS3FullAccess policies to your existing IAM role, which is more permissive. For more information about the AmazonSageMakerFullAccess policy, see AmazonSageMakerFullAccess.

If you wish to attach the existing policies to your IAM role, you may skip the instructions set here and continue following the instructions under To add permissions to your IAM role.

The following instructions creates a custom IAM policy that is tailored to this service with minimum permissions.

Open the Amazon SageMaker console at https://console.aws.amazon.com/sagemaker/.
In the search bar at the top of the page, enter IAM.
Under Services, select Identity and Access Management (IAM).
Choose Policies from the navigation pane.
Choose Create policy. When the Policy editor opens, choose JSON.

Ensure that the following permissions appear in the Policy editor. You can also copy and paste the following into the Policy editor.


{
"Version": "2012-10-17",
"Statement":
    [{
        "Effect": "Allow",
        "Action": [
            "s3:GetObject",
            "s3:PutObject",
            "s3:ListBucket"
        ],
        "Resource": [
            "arn:aws:s3:::{input_bucket}/*",
            "arn:aws:s3:::{input_bucket}",
            "arn:aws:s3:::{output_bucket}/*",
            "arn:aws:s3:::{output_bucket}",
            "arn:aws:s3:::jumpstart-cache-prod-{region}/*",
            "arn:aws:s3:::jumpstart-cache-prod-{region}"
        ]
    },
    {
        "Effect": "Allow",
        "Action": [
            "sagemaker:CreateEndpoint",
            "sagemaker:DeleteEndpoint",
            "sagemaker:CreateEndpointConfig",
            "sagemaker:DeleteEndpointConfig"
        ],
        "Resource": [
            "arn:aws:sagemaker:{region}:{account-id}:endpoint/sm-margaret-*",
            "arn:aws:sagemaker:{region}:{account-id}:endpoint-config/sm-margaret-*"
        ],
        "Condition": {
            "ForAnyValue:StringEquals": {
                "aws:TagKeys": "sagemaker-sdk:jumpstart-model-id"
            }
        }
    },
    {
        "Effect": "Allow",
        "Action": [
            "sagemaker:DescribeProcessingJob",
            "sagemaker:DescribeEndpoint",
            "sagemaker:InvokeEndpoint"
        ],
        "Resource": "*"
    },
    {
        "Effect": "Allow",
        "Action": [
            "sagemaker:DescribeInferenceComponent",
            "sagemaker:AddTags",
            "sagemaker:CreateModel",
            "sagemaker:DeleteModel"
        ],
        "Resource": "arn:aws:sagemaker:{region}:{account-id}:model/*",
        "Condition": {
            "ForAnyValue:StringEquals": {
                "aws:TagKeys": "sagemaker-sdk:jumpstart-model-id"
            }
        }
    },
    {
        "Effect": "Allow",
        "Action": [
            "sagemaker:DescribeFlowDefinition",
            "sagemaker:StartHumanLoop",
            "sagemaker:DescribeHumanLoop"
        ],
        "Resource": "*"
    },
    {
        "Effect": "Allow",
        "Action": [
            "logs:CreateLogStream",
            "logs:PutLogEvents",
            "logs:CreateLogGroup",
            "logs:DescribeLogStreams"
        ],
        "Resource": "arn:aws:logs:{region}:{account-id}:log-group:/aws/sagemaker/ProcessingJobs:*"
    },
    {
        "Effect": "Allow",
        "Action": [
            "cloudwatch:PutMetricData"
        ],
        "Resource":"*"
    },
    {
        "Effect": "Allow",
        "Action": [
            "ecr:GetAuthorizationToken",
            "ecr:BatchCheckLayerAvailability",
            "ecr:GetDownloadUrlForLayer",
            "ecr:BatchGetImage"
        ],
        "Resource": "*"
    },
    {
        "Effect": "Allow",
        "Action": [
            "kms:DescribeKey",
            "kms:GetPublicKey",
            "kms:Decrypt",
            "kms:Encrypt"
        ],
        "Resource": [
            "arn:aws:kms:{region}:{account-id}:key/{kms-key-id}"
        ]
    },
    {
        "Effect": "Allow",
        "Action": [
            "iam:PassRole"
        ],
        "Resource": "arn:aws:iam::{account-id}:role/{this-role-created-by-customer}",
        "Condition": {
            "StringEquals": {
                "aws:PrincipalAccount": [
                    "account-id"
                ]
            }
        }
    }]
}

Choose Next.
Enter a policy name in the Policy details section, under Policy name. You can also enter an optional description. You will search for this policy name when you assign it to a role.
Choose Create policy.

To add permissions to your IAM role

Open the Amazon SageMaker console at https://console.aws.amazon.com/sagemaker/.
In the search bar at the top of the page, enter IAM.
Under Services, select Identity and Access Management (IAM).
Choose Roles in the navigation pane.
If you are creating a new role:
1. Choose Create role.
2. On the Select trusted entity step, under Trusted entity type choose Custom trust policy.
3. In the Custom trust policy editor, next to Add principal choose Add.
4. On the Add principal pop-up box, under Principal type select AWS services from the dropdown list of options.
5. Under ARN replace {ServiceName} with sagemaker.
6. Choose Add principal.
7. Choose Next.
8. (Optional) Under Permissions policies select the policies you would like to add to your role.
9. (Optional) Under Set permissions boundary - optional choose your permission boundary setting.
10. Choose Next.
11. On the Name, review, and create step, under Role details fill in your Role name and Description.
12. (Optional) Under Add tags - optional, you can add tags by choosing Add new tag and enter a Key and Value - optional pair.
13. Review your settings.
14. Choose Create role.
If you are adding the policy to an existing role:
1. Select the name of the role under Role name. The main window changes to show information about your role.
2. In the Permissions policies section, choose the down arrow next to Add permissions.
3. From the options that appear, choose Attach policies.
4. From the list of policies that appear, search for and select the policy that you created under To add permissions to your IAM policy and select the check the box next to your policy's name. If you did not create a custom IAM policy, search for and select the check boxes next to the AWS provided AmazonSageMakerFullAccess and AmazonS3FullAccess policies. You may want to consider the level of permissions to attach to your IAM role. The instructions for the custom IAM policy is less permissive, while the latter is more permissive. For more information about the AmazonSageMakerFullAccess policy, see AmazonSageMakerFullAccess.
5. Choose Add permissions. A banner at the top of the page should state Policy was successfully attached to role. when completed.

To add trust policy to your IAM role

The following trust policy makes it so administrators can allow SageMaker to assume the role. You need to add the policy to your IAM role. Use the following steps to do so.

Open the Amazon SageMaker console at https://console.aws.amazon.com/sagemaker/.
In the search bar at the top of the page, enter IAM.
Under Services, select Identity and Access Management (IAM).
Choose Roles in the navigation pane.
Select the name of the role under Role name. The main window changes to show information about your role.
Choose the Trust relationship tab.
Choose Edit trust policy.

Ensure that the following policy appears under Edit trust policy. You can also copy and paste the following into the editor.


{
"Version": "2012-10-17",
"Statement": [
    {
        "Sid": "",
        "Effect": "Allow",
        "Principal": {
            "Service": [
                "sagemaker.amazonaws.com"
            ]
        },
        "Action": "sts:AssumeRole"
    }
]
}

Choose Update policy. A banner at the top of the page should state Trust policy updated. when completed.

You can create a human evaluation job using a text-based model that is available in JumpStart or you can use a JumpStart model that you've previously deployed to an endpoint.

To launch JumpStart

Open the Amazon SageMaker console at https://console.aws.amazon.com/sagemaker/.
In the search bar at the top of the page, enter SageMaker.
Under Services, select Amazon SageMaker.
Choose Studio from the navigation pane.
Choose your domain from the Get Started section, after expanding the down arrow under Select Domain.
Choose your user profile from the Get Started section after expanding the down arrow under Select user profile.
Choose Open Studio to open the landing page for Studio.
Choose Jobsfrom the navigation pane.

To set up an evaluation job

On the Model evaluation home page, choose Evaluate a model
Specify job details.
1. Enter the Evaluation name of your model evaluation. This name helps you identify your model evaluation job after it is submitted.
2. Enter a Description to add more context to the name.
3. Choose Next.
Set up evaluation
1. Under Choose an evaluation type, select the radio button next to Human.
2. Under Choose the model(s) you want to evaluate, choose Add model to evaluation. You can evaluate up to two models for each evaluation.
  1. To use a pre-trained JumpStart model, choose Pre-trained JumpStart foundation model. If you want to use a JumpStart model that you have previously deployed to an endpoint, choose Endpoints with JumpStart foundation models.
  2. If the model requires a legal agreement, select the check box to confirm that you agree.
  3. If you want to add another model, repeat the previous step.
3. To change how the model behave during inference choose, Set parameters.
  
  Set parameters contains a list of inference parameters that affect the degree of randomness in your model's output, the length of your model's output, and what words the model will choose next.
4. Next, select an Task type. You can select any of the following:
  - Text Summarization
  - Question Answering (Q&A)
  - Text classification
  - Open-ended Generation
  - Custom
5. In the Evaluation metrics section, choose an Evaluation dimension and enter additional context about the dimension in the text box under Description. You can choose from the following dimensions:
  - Fluency – Measures the linguistic quality of a generated text.
  - Coherence – Measures the organization and structure of a generated text.
  - Toxicity – Measures the harmfulness of a generated text.
  - Accuracy– Indicates the accuracy of a generated text.
  - A custom evaluation dimension that you can define the name and description of for your work team.
    
    To add a custom evaluation dimension, do the following:
    Choose Add an evaluation dimension.
    
    In the text box containing Provide evaluation dimension, input the name of your custom dimension.
    
    In the text box containing Provide description for this evaluation dimension, input a description so that your work team understands how to evaluate your custom dimension.
  Under each of these metrics are reporting metrics that you can choose from the Choose a metric type down arrow. If you have two models to evaluate, you can choose either comparative or individual reporting metrics. If you have one model to evaluate, you can choose only individual reporting metrics. You can choose the following reporting metrics types for each of the above metrics.
  - (Comparative) Likert scale - comparison – A human evaluator will indicate their preference between two responses on a 5-point Likert scale according to your instructions. The results in the final report will be shown as a histogram of preference strength ratings from the evaluators over your whole dataset. Define the important points of the 5-point scale in your instructions so that your evaluators know how to rate the responses according to your expectations. In the JSON output saved in Amazon S3 this choice is represented as ComparisonLikertScale the key value pair "evaluationResults":"ComparisonLikertScale".
  - (Comparative) Choice buttons – Allows a human evaluator to indicate their one preferred response over another response. Evaluators indicate their preference between two responses according to your instructions using radio buttons. The results in the final report will be shown as a percentage of responses that workers preferred for each model. Explain your evaluation method clearly in your instructions. In the JSON output saved in Amazon S3 this choice is represented as ComparisonChoice the key value pair "evaluationResults":"ComparisonChoice".
  - (Comparative) Ordinal Rank – Allows a human evaluator to rank their preferred responses to a prompt in order, starting at 1, according to your instructions. The results in the final report will be shown as a histogram of the rankings from the evaluators over the whole dataset. Define the what a rank of 1 means in your instructions. In the JSON output saved in Amazon S3 this choice is represented as ComparisonRank the key value pair "evaluationResults":"ComparisonRank".
  - (Individual) Thumbs up/down – Allows a human evaluator to rate each response from a model as acceptable or unacceptable according to your instructions. The results in the final report will be shown as a percentage of the total number of ratings by evaluators that received a thumbs up rating for each model. You may use this rating method for an evaluation one or more models. If you use this in an evaluation that contains two models, a thumbs up or down will be presented to your work team for each model response and the final report will show the aggregated results for each model individually. Define what is acceptable as a thumbs up or thumbs down rating in your instructions. In the JSON output saved in Amazon S3 this choice is represented as ThumbsUpDown the key value pair "evaluationResults":"ThumbsUpDown".
  - (Individual) Likert scale - individual – Allows a human evaluator to indicate how strongly they approve of the model response based on your instructions on a 5-point Likert scale. The results in the final report will be shown as a histogram of the 5-point ratings from the evaluators over your whole dataset. You may use this scale for an evaluation containing one or more models. If you select this rating method in an evaluation that contains more than one model, a 5-point Likert scale will be presented to your work team for each model response and the final report will show the aggregated results for each model individually. Define the important points on the 5-point scale in your instructions so that your evaluators know how to rate the responses according to your expectations. In the JSON output saved in Amazon S3 this choice is represented as IndividualLikertScale the key value pair "evaluationResults":"IndividualLikertScale".
6. Choose a Prompt dataset. This dataset is required and will be used by your human work team to evaluate responses from your model. Provide the S3 URI to an Amazon S3 bucket that contains your prompt dataset in the text box under S3 URI for your input dataset file. Your dataset must be in jsonlines format and contain the following keys to identify which parts of your dataset the UI will use to evaluate your model:
  - prompt – The request that you want your model to generate a response to.
  - (Optional) category – - The category labels for your prompt. The category key is used to categorize your prompts so you can filter your evaluation results later by category for a deeper understanding of the evaluation results. It does not participate in the evaluation itself, and workers do not see it on the evaluation UI.
  - (Optional) referenceResponse – The reference answer for your human evaluators. The reference answer is not rated by your workers, but can be used to understand what responses are acceptable or unacceptable, based on your instructions.
  - (Optional) responses – Used to specify inferences from a model outside of SageMaker or outside of AWS.
    
    This object requires two additional key value pairs "modelIdentifier which is a string that identifies the model, and "text" which is the model's inference.
    
    If you specify a "responses" key in any input of the of custom prompt dataset it must be specified in all inputs.
  - The following json code example shows the accepted key-value pairs in a custom prompt dataset. The Bring your own inference check box must be checked if a responses key is provided. If checked, the responses key must always be specified in each prompt. The following example could be used in a question and answer scenario.
```
{
    "prompt": {
        "text": "Aurillac is the capital of"
    },
    "category": "Capitals",
    "referenceResponse": {
        "text": "Cantal"
    },
    "responses": [
        // All responses must come from a single model. If specified it must be present in all JSON objects. modelIdentifier and text are then also required.
        {
            "modelIdentifier": "meta-textgeneration-llama-codellama-7b",
            "text": "The capital of Aurillac is Cantal."
        }
    ]
}
```
7. Input an S3 bucket location where you want to save the output evaluation results in the text box under Choose an S3 location to save your evaluation results. The output file written to this S3 location will be in JSON format, ending in the extension,.json.
8. Note
  If you want to include bring your own inference data in the model evaluation job, you can only use a single model.
  
  (Optional) Choose the check box under Bring your own inference to indicate that your prompt dataset contains the responses key. If you specify the responses key as part of any prompts it must be present in all of them.
9. Configure your processor in the Processor configuration section using the following parameters:
  - Use Instance count to specify the number of compute instances to use to run your model. If you use more than 1 instance, your model will run in parallel instances.
  - Use Instance type to choose the kind of compute instance you want to use to run your model. AWS has general compute instances and instances that are optimized for computing and memory. For more information about instance types, see Instance types available for use with Studio Classic .
  - If you want SageMaker to use your own AWS Key Management Service (AWS KMS) encryption key instead of the default AWS managed service key, toggle to select On under Volume KMS key, and input the AWS KMS key. SageMaker will use your AWS KMS key to encrypt data on the storage volume. For more information about keys, see AWS Key Management Service.
  - If you want SageMaker to use your own AWS Key Management Service (AWS KMS) encryption key instead of the default AWS managed service key, toggle to select On under Output KMS key and input the AWS KMS key. SageMaker will use your AWS KMS key to encrypt the processing job output.
  - Use an IAM role to specify the access and permissions for the default processor. Input the IAM role that you set up in the section Set up your IAM role in this Run a human evaluation section.
10. After you specify your model and criteria, select Next.

Your work team consists of the people that are evaluating your model. After your work team is created, it persists indefinitely and you cannot change its attributes. The following shows how to get started with your work team.

Set up your work team

Choose an existing team or Create a new team in the Select team input text box.
Specify a name of your organization in Organization name. This field only appears when you create the first work team in the account.
Specify a contact email. Your workers will use this email to communicate with you about the evaluation task that you will provide to them. This field only appears when you create the first work team in the account.
Specify a Team name. You cannot change this name later.
Specify a list of Email addresses for each of your human workers that will evaluate your large language model (LLM). When you specify the email addresses for your team, they are notified of a new job only when they are newly added to a work team. If you use the same team for a subsequent job, you must notify them manually.
Then, specify the Number of workers per prompt

Provide instructions for your work team

Provide detailed instructions to your human workforce so that they can evaluate your model to your metrics and standards. A template in the main window shows sample instructions that you can provide. For more information about how to give instructions, see Creating good worker instructions.
To minimize bias in your human evaluation, select the check box next to Randomize response positions.
Select Next.

You can review the summary of the selections that you have made for your human job. If you must change your job, choose Previous to go back to an earlier selection.

Submit your evaluation job request and view job progress

To submit your evaluation job request, choose Create resource.
To see the status of all of your jobs, choose Jobs in the navigation pane. Then, choose Model evaluation. The evaluation status displays as Completed, Failed, or In progress.

The following also displays:
- Sample notebooks to run a model evaluation in SageMaker and Amazon Bedrock.
- Links to additional information including documentation, videos, news, and blogs about the model evaluation process.
- The URL to your Private worker portal is also available.
Select your model evaluation under Name to view a summary of your evaluation.
- The summary gives information about the status of the job, what kind of evaluation task you ran on which model, and when it ran. Following the summary, the human evaluation scores are sorted and summarized by metric.

View the report card of you model evaluation job that uses human workers

To see the report for your jobs, choose Jobs in the navigation pane.
Then, choose Model evaluation. One the Model evaluations home page, use the table to find your model evaluation job. Once the job status has changed to Completed you can view your report card.
Choose the name of the model evaluation job to it's report card.

When you create a model evaluation job that uses human workers you have the option to bring your own inference data, and have your human workers compare that inference data to data produced by one other JumpStart model or a JumpStart model that you have deployed to an endpoint.

This topic describes the format required for the inference data, and a simplified procedure for how to add that data to your model evaluation job.

Choose a Prompt dataset. This dataset is required and will be used by your human work team to evaluate responses from your model. Provide the S3 URI to an Amazon S3 bucket that contains your prompt dataset in the text box under Choose an S3 location to save your evaluation results. Your dataset must be in .jsonl format. Each record must be a valid JSON object, and contain the following required keys:

prompt – A JSON object that contains the text to be passed into the model.
(Optional) category – - The category labels for your prompt. The category key is used to categorize your prompts so you can filter your evaluation results later by category for a deeper understanding of the evaluation results. It does not participate in the evaluation itself, and workers do not see it on the evaluation UI.
(Optional) referenceResponse – a JSON object that contains the reference answer for your human evaluators. The reference answer is not rated by your workers, but can be used to understand what responses are acceptable or unacceptable, based on your instructions.
responses – Used to specify individual inferences from a model outside of SageMaker or outside of AWS.

This object requires to additional key value pairs "modelIdentifier which is a string that identifies the model, and "text" which is the model's inference.

If you specify a "responses" key in any input of the of custom prompt dataset it must be specified in all inputs.

The following json code example shows the accepted key-value pairs in a custom prompt dataset that contains your own inference data.


{
    "prompt": {
        "text": "Who invented the airplane?"
    },
    "category": "Airplanes",
    "referenceResponse": {
        "text": "Orville and Wilbur Wright"
    },
    "responses":
        // All inference must come from a single model
        [{
            "modelIdentifier": "meta-textgeneration-llama-codellama-7b" ,
            "text": "The Wright brothers, Orville and Wilbur Wright are widely credited with inventing and manufacturing the world's first successful airplane."
        }]

}

To get started launch Studio, and under choose Model evaluation under Jobs in the primary navigation.

To add your own inference data to a human model evaluation job.

In Step 1: Specify job details add the name of your model evaluation job, and an optional description.
In Step 2: Set up evaluation choose Human.
Next, under Choose the model(s) you want to evaluate you can choose the model that you want to use. You can use either a JumpStart model that has already deployed or you can choose a Pre-trained Jumpstart foundation model.
Then, choose a Task type.
Next, you can add Evaluation metrics.
Next, under Prompt dataset choose the check box under Bring your own inference to indicate that your prompts have response keys in it.
Then continue setting up your model evaluation job.

To learn more about how the responses from your model evaluation job that uses human workers are saved, see Understand the results of a human evaluation job

Warning Javascript is disabled or is unavailable in your browser.

To use the Amazon Web Services Documentation, Javascript must be enabled. Please refer to your browser's Help pages for instructions.

Document Conventions

Toxicity

Automatic model evaluation