This repository provides Google Colab notebooks for experimenting with different guardrail frameworks used in securing and controlling Large Language Models (LLMs).
It covers hands-on tutorials, code samples, and comparison experiments across three widely discussed frameworks:
- NVIDIA NeMo Guardrails
- Llama Guard (Meta’s safety classifier)
- Guardrails AI
- Rule-based + ML-driven framework for enforcing safety, security, and compliance.
- Supports flow definition files (Colang) for conversational control.
- Integration with multiple LLM providers.
- Example use cases: content filtering, data redaction, safety checks.
- Lightweight guardrail model released by Meta.
- Functions as a safety classifier that detects policy-violating generations.
- Can be used as a filter alongside other LLMs.
- Example use cases: toxicity detection, harmful instruction blocking.
- Open-source Python library for specifying, validating, and enforcing constraints on LLM outputs.
- Schema-driven approach with pydantic-like validators.
- Strong focus on structured outputs (JSON/XML/YAML) and semantic validation.
- Example use cases: ensuring responses follow format, preventing hallucinations.
Each notebook can be opened and run directly in Google Colab
- Ensure you have access to GPU runtime in Colab (
Runtime > Change runtime type > GPU). - Some frameworks may require API keys or model downloads (instructions included in each notebook).
This repository is released under the MIT License.