Skip to main content
CAIL Sandbox Docs

Custom Models

Custom models combine a base model with instructions, documents, and tools for tasks such as reviewing drafts or working through course readings. You can specify which materials the model should reference, how it should respond, and which requests it should decline. Creating this configuration does not train a new base model.

Creating a Model

  1. Request Workspace access by emailing the CUNY AI Lab team. Once enabled, click Workspace in the left sidebar.
  2. Select Models
  3. Click Create beside the Workspace tabs
    • Enter a Model Name. In the editor, define its behavior, connect documents, and choose its tools.
  4. Choose a base model from the dropdown
    • Test the available models with the kinds of prompts you plan to use before choosing one. If you need a specific model and do not see it listed, contact the CUNY AI Lab team.
  5. Write a system prompt (see below for guidance)
  6. Attach knowledge bases if you want the model to draw from your documents
  7. Select Tools for connected services, and configure Web Search or Code Interpreter under Capabilities and Default Features if available
  8. Add prompt suggestions
    • These appear as clickable suggestions above the input bar when students open a new chat. Write examples of tasks or questions the model is configured to support.
  9. Open Access
    • Keep Private while building. Use Add Access to share with your course group or selected users.
    • Choose Public only if the model should be available to all signed-in Sandbox users and your account permits it.
  10. Click Save & Create (or Save & Update when editing)

System Prompt

In step 5, write a system prompt that specifies the model’s role, boundaries, and instructional approach. For patterns, examples, and advanced prompt techniques, see System Prompts as Instructional Design.

Prompt Suggestions

Write short prompts that students can select to begin a task, such as choosing a statistical test or interpreting a result. Examples for a data analysis model

  • “What visualization works best for this data?”
  • “Help me interpret these statistical results”
  • “Which test should I use for this research question?”

Advanced Settings

View details

Advanced Parameters

These parameters control response generation; adjust them when the default settings do not fit your task.

  • Max Tokens caps output length. One token is roughly three-quarters of an English word. Lower limits (100-500) can cut responses short; higher limits (1000-4000) leave more room. Use the prompt to request concise responses.
  • Temperature controls randomness. Lower values (0.1 to 0.3) reduce variation; higher values (0.7 to 1.0) allow more variation. Low temperature does not guarantee identical or accurate responses.
  • Top P (nucleus sampling) controls diversity of word selection. Leave at default unless you have a specific reason to change it.
  • Stop Sequences force the model to stop generating when it encounters specific text strings. Enter sequences like <|end_of_text|> or User: and press Enter.

Switching Models Mid-Chat

You can change models during a chat by clicking the model name on the right inside the message box. The chat context carries across the switch. This lets you use different models for different stages of a task.

Tip. Ask students to compare how two models respond to the same prompt.

Callout

For instructors. Specify which parts of an assignment the model should support and what it should do when students ask for work outside those limits. Test the configuration with likely student requests, then explain the permitted uses in the assignment instructions.

Additional Resources

← Return to Getting Started | Continue to Knowledge Bases →