TritonGPT Guardrails

This repository contains guardrail configuration and classifier prompts used by TritonGPT deployments.

Files

• generic_safety_prompts.yaml - Generic safety guardrail configuration.
• ucsd_safety_prompts.yaml - UC San Diego guardrail configuration.
• berkeley_guardrail.yml - UC Berkeley guardrail configuration.
• prompts/agent-selfharm.yml - Self-harm classifier prompt.
• prompts/agent-unsafe.yml - General unsafe-content classifier prompt.
• prompts/agent-unions.yml - Union-content classifier prompt.

The root guardrail files use the split-agent format:

version: "2.0.0"

agents:
  - name: "selfharm"
    priority: 0
    label_name: "SELF_HARM"
    prompt_url: "https://raw.githubusercontent.com/ucsd-ets/tritongpt-guardrails/main/prompts/agent-selfharm.yml"
    response: |
      ...

Each agent points to a prompt file and defines the response returned when that classifier flags a request.