​
Overview

The goal of this guide is to help project leads (Admins, Managers, Members) navigate the user interface of the Scale Generative AI platform. We are making rapid improvements to the platform and will keep this page updated as we release new features.

​
Admin Dashboard

An account is simply an organizational unit that is used to isolate resources from one another in Scale GP. It can be used to represent a team, a use case, or simply a namespace for a specific project. To view accounts, click “Admin” on the left side panel, and click the “Accounts” tab.

Below are several actions you may want to perform and instructions on how to do them.

​
Register a New User

All new users are registered with Scale GP automatically by your company’s SAML provider when they first visit the Scale GP app. If a user would like to gain access to the application, simply navigate to the homepage.

The homepage URL will have the following syntax. Your Scale GP representative will have given this URL to an administrator at your company. Please reach out to them for the full URL.

https://app.<your_workspace_id>.workspace.egp.scale.com

Multi-Tenant Scale GP UsersIf you are demoing Scale GP using the multi-tenant Scale-hosted version of our product, the url is https://gp.scale.com and your username / login information will have been given to you by a Scale representative.

​
Adding a user to a specific account

Prerequisites- The user must have first registered by going to the app themselves- The person adding the user to the account must be an “admin” or “manager” of the account.

  • Navigate to the “Accounts” tab

  • Click an account

  • Click the “Add User” button and enter the user’s email address

  • A success banner should appear saying the user was successfully added

Noteif an error occurs to add a user to an account, it’s likely that they did not first access the site themselves. Please instruct them to do so.

​
Change a users’ role

  • Navigate to the “Accounts” tab

  • Click an account

  • Update the user’s role in the “role column”

​
Role Definitions

These are the permissions we will grant to each of the account level roles. Note that this will likely change in the future and we can update the doc accordingly.

​
Admin/Manager

  • Add user to account / approve access requests

  • Change a user’s role for an account

  • Create, view, update, and delete SGP resources (applications, evaluations, datasets, models, etc.)

  • View and install resources (models, etc.) from component catalog

  • Audit evaluation tasks

  • Manage contributors and evaluation analytics

  • All permissions of Editor, Members and Labelers

​
Editor

  • Create, view, update, and delete SGP resources (applications, evaluations, datasets, models, etc.)

  • View and install resources (models, etc.) from component catalog

  • Audit evaluation tasks

  • Manage contributors and evaluation analytics

  • All permissions of Members and Labelers

​
Member

  • View/Use existing SGP resources (applications, evaluations, datasets, models, etc.)

  • View and rotate one’s own API key

  • All permissions of Labelers

​
Labeler

  • View and evaluate tasks

  • Request access to an account

  • Annotations Dashboard (Includes Auditing Instructions)

NoteThe views for evaluations will look different for contributors versus admins/managers. To learn more about what the experience looks like for contributors, please refer to the contributors quick guide. This doc will focus on the admin/manager experience.

​
Evaluation > Evaluation Runs

An evaluation run combines a set of test cases with the outputs for these test cases of a GenAI application and sends these off as a batch of tasks to be evaluated based on a set of evaluation criteria. You can create a new evaluation by clicking the “New Evaluation” button.

URL: `https://app.<your_workspace_id>.workspace.egp.scale.com/evaluation/runs

The above page contains a list of evaluations run by developers on the Scale GP platform. Every time a developer submits evaluations using the REST API or SDK, a table row will show up that shows the AI output alongside the evaluation dataset input that was used the generate the output.

If you click into a specific evaluation run, you’ll be able to view test case results. A test case is the individual input (e.g. “what is the price target for apple”) that is being tested for a given Gen AI application. The Gen AI app produces an output for the input and this is then evaluated based on things like accuracy, coherence etc. Together, the input, output and evaluation are the test case result.

​
View Contributor Progress and Audit Evaluations

Click the “Annotations” tab on the left sidebar. If there are multiple projects available, they will all appear in this view. Users can start labeling a specific project by clicking on “Start Labeling.” Under the Contributor Dashboard, you will see two cards:

  • Evaluations in Progress

  • Evaluations Completed

Under the Evaluations in Progress Card, you will be able to view evaluation progress (# completed / # total evaluations). There will also be two buttons:

  • Task Dashboard

  • Start Labeling

If you click the “Task Dashboard” button, you will be able to view some metrics, including:

  • Project throughput status (# completed / # total evaluations)

  • # Tasks flagged for review

  • Audit statistics (You can use this as a preliminary indication of quality)

    • % of tasks unaudited, accepted and fixed by the auditing team.

  • Average time per task

You can also view a deep-dive table which includes additional information about your evaluation that may be useful including when the task was last updated, statuses, how long it took to complete, who it was assigned to, when it was assigned etc. You’ll also see an actions column which includes an option to View, Audit, or Unassign a task.

Admins and managers can unassign evaluations but this functionality will be disabled for contributors. Unassignment needs to be done only if a contributor has claimed a task, did not complete it, went on leave, and is now blocking the completion of the task queue because all other tasks have been completed. An account admin or manager can then access the task dashboard, un-assign the task, and any contributor can then pick up the unlocked task to complete it.

If the task is not yet completed, you will see a “View” button, which will surface a read-only version of the evaluation.

If the task is completed, you’ll see an “Audit” button.

After you click “Audit,” you will be able to view the completed task.

If you click “Fix,” you’ll be able to change contributor responses. After you fix the task, click “Submit” to save your audit.