Troubleshooting¶
This page focuses on the issues most likely to block a first OSS user: missing extras, auth problems, path mistakes, and backend expectations that are easy to miss.
geh Command Not Found¶
Make sure you installed the project in the environment you are actively using:
If that still fails, activate the right virtual environment and try again.
Adapter Exists In Docs But Not In geh list backends¶
This usually means the required extra is missing.
Examples:
hf,hf_vlm_guardneedpip install -e ".[hf]"vllmneedspip install -e ".[vllm]"openai_moderation,openai_compatible,anthropic, andhttpneedpip install -e ".[api]"
OpenAI, Anthropic, Or Gated HuggingFace Auth Fails¶
Check the environment variables first:
Common fixes:
- export the key in the current shell before running
geh - copy
.env.exampleto.envand source it in your shell tooling - make sure the selected model or dataset is actually accessible to your account
geh validate --config ... Fails On A Local Path¶
Local dataset paths in YAML are resolved relative to the config file location, not the shell directory you happened to run the command from.
If a config lives in examples/, use paths that make sense relative to that
file.
Pack Name Confusion¶
These are both valid:
The stable shorthand resolves to the current versioned pack alias.
Report Or Export Steps Fail After A Run¶
The safest way to debug is to inspect the run directory first:
Confirm that these files exist:
manifest.jsonsummary.jsondatasets/<dataset>/metrics.json
Local Multimodal Runs Fail On File Paths¶
For local_image_jsonl and local_image_dir:
- verify the path exists on disk
- use absolute paths if you are debugging path resolution
- confirm the manifest fields point to files the current machine can read
GPU Backends OOM Or Run Too Slowly¶
Start by reducing batch size. For local text models, prefer:
For multimodal adapters, start conservatively with batch size 1 or 2.
A Run Stops Partway Through¶
Use resume mode for long-running jobs:
Then rerun the same config with the same output.run_dir.
Need A Known-Good Starting Point¶
Use one of these first:
geh run --dataset xstest --model mock --limit 50geh run --config examples/run-mock-jsonl.yamlgeh run --pack core --model openai_moderation --limit 100