Installation¶
Use the base install when you only need the CLI, configs, dataset normalization, and core metrics. Add extras only for the backends and features you plan to use.
Requirements¶
- Python
3.10+ pipor another PEP 517-compatible installer- a working shell environment for API keys if you use hosted models
Base Install¶
git clone https://github.com/Virtue-Research/guard-eval-harness.git
cd guard-eval-harness
pip install -e "."
The base install is enough for:
- the
gehCLI - config validation
- built-in dataset loading
- the
mockmodel adapter - artifact inspection and comparison flows
Optional Extras¶
Use this for local text, image, and specialized multimodal HuggingFace adapters.
Useful when you want HTML or spreadsheet-friendly reporting dependencies available explicitly.
Environment Setup¶
Copy the example environment file if you plan to use hosted models or gated HuggingFace assets:
You only need to set the variables required by the adapters you actually run.
Verify The Install¶
You should see JSON output with the created run_dir, manifest_path, and
summary_path.
Next Steps¶
- Quickstart for a 2-minute first run
- Run Modes to choose inline, pack, or YAML config flows
- Troubleshooting if installs, auth, or dataset downloads fail