CLI Reference¶
Guard Eval Harness exposes a single geh CLI for running evaluations,
inspecting artifacts, and discovering available adapters and suites.
Global Usage¶
Most commands print JSON so the output is easy to inspect manually or pipe into other tooling.
geh run¶
Run a benchmark and write artifacts to disk.
Source Flags¶
| Flag | Meaning |
|---|---|
--config PATH |
run from a YAML configuration file |
--pack NAME |
run a curated benchmark pack such as core |
--dataset NAMES |
run one or more comma-separated dataset adapters inline |
Model Flags¶
These are required when using --pack or --dataset.
| Flag | Meaning |
|---|---|
--model ADAPTER |
model adapter alias, and for pack flows may also resolve a supported catalog slug |
--model-name NAME |
backend-specific model identifier |
--model-args JSON |
adapter-specific JSON args |
Common Options¶
| Flag | Default | Meaning |
|---|---|---|
--batch-size INT |
1 |
batch size for inference |
--threshold FLOAT |
0.5 |
unsafe classification threshold |
--limit INT |
all | max samples per dataset |
--output-dir PATH |
auto | override artifact output path |
--no-sample-cache |
false |
disable image/media sample caching |
Examples¶
geh run --dataset xstest --model mock --limit 50
geh run --dataset xstest,toxic_chat --model hf --model-name meta-llama/Llama-Guard-3-8B
geh run --pack core --model openai_moderation --limit 100
geh run --config examples/run-mock-jsonl.yaml
geh list¶
Discover registered components.
What Each Target Means¶
| Target | What it returns |
|---|---|
datasets |
registered dataset adapter aliases |
backends |
registered model adapter aliases |
metrics |
built-in metric names |
packs |
curated user-facing benchmark packs |
presets |
code-defined reproducible benchmark suites, currently including 21x31 |
plugins |
the active dataset/model registry view after built-ins and entry-point plugins are loaded |
Examples¶
geh list plugins is useful when you want to confirm that an installed
entry-point plugin was actually discovered by the harness.
geh inspect¶
Inspect the stored manifest, summary, and dataset artifact layout for a prior run.
geh report¶
Rebuild summary artifacts from an existing run directory.
Use this when the run already exists and you want to regenerate summary.json
or report.html from stored outputs.
geh compare¶
Compare two run directories.
Useful for model swaps, threshold changes, or regressions across the same dataset mix.
geh export¶
Export summary artifacts to a simpler stakeholder format.
geh validate¶
Validate a YAML config without running the full benchmark.
This resolves adapters, validates local dataset paths, and confirms the config can be materialized into a run.
geh cache¶
Inspect or clear the normalized sample cache used for media-heavy datasets.
Examples: