Quickstart¶
This guide gets you from a fresh checkout to a finished evaluation run quickly, then shows the cleanest next step for a real backend.
1. Install The Base Package¶
2. Run A Smoke Test¶
Why this first:
- no GPU required
- no API keys required
- you get the full artifact layout immediately
3. Inspect The Output¶
The command prints a JSON payload with the run directory and artifact paths. The output directory looks like this:
out/mock/xstest/
manifest.json
resolved-config.json
summary.json
report.html
datasets/
xstest/
predictions.jsonl
metrics.json
dataset-manifest.json
Follow it with:
4. Try A Real Backend¶
Local HuggingFace¶
pip install -e ".[hf]"
geh run --dataset xstest,toxic_chat \
--model hf \
--model-name meta-llama/Llama-Guard-3-8B \
--batch-size 16
OpenAI Moderation¶
pip install -e ".[api]"
export OPENAI_API_KEY=sk-...
geh run --pack core --model openai_moderation --limit 100
Reproducible YAML Run¶
5. Compare Or Export Results¶
geh compare --run-a out/run-a --run-b out/run-b
geh export --run-dir out/mock/xstest --format csv --output results.csv
What To Do Next¶
- Use Run Modes if you are not sure when to choose inline, pack, or YAML config runs.
- Use Benchmark Selection to find the right dataset mix.
- Use Common Workflows for copy-paste recipes across text and image evaluation.