obarlik's picture
Upload README.md with huggingface_hub
f137f9c verified
|
Raw History Blame Contribute Delete
2.74 kB
---
license: apache-2.0
task_categories: [text-classification]
tags: [agents, agent-evaluation, llm-as-a-judge, benchmark, process-quality]
pretty_name: Journeyman judge-calibration set
---
# Journeyman — judge-calibration set (82 cases, 7 axes)
This is the exam a judge sits before it is allowed to score anything in
[Journeyman](https://github.com/codechu/journeyman), a process-quality
benchmark for LLM agents. Each case is a real agent record plus the label a
judge should return for one axis. A judge that misses any axis does not get
the badge; the ones that failed are published too, in the
[judge registry](https://github.com/codechu/journeyman/blob/HEAD/docs/judges.md).
## What is in it
| axis | cases |
|---|---|
| `empty-measure` | 17 |
| `grounding` | 14 |
| `handoff-verification` | 11 |
| `object-hold` | 8 |
| `relief-page` | 9 |
| `route-discipline` | 9 |
| `wall-pricing` | 14 |
Labels come from a three-family council (claude-sonnet-5, kimi-k2, grok-4.3):
a blind round, then an anonymous round where each labeller had to quote its
evidence. A label is sealed only with support from at least two families; the
maintainer ruled the cases the council split on; empty-measure is counted
mechanically under its published definition.
## Honest limits
- **82 cases.** Small next to human-annotated process benchmarks. It is sized
for what it claims — licensing a judge — not for ranking a field.
- **Labels are model-produced**, not human. The council is cross-family and the
protocol was frozen before labelling, but this is not a human gold standard.
- **The set has been corrected**, and every correction is in the file's own
`note` field: four cases whose only closing report the scene refused were
relabelled, twelve fresh cases were harvested when unfiled reports came to
dominate two axes, and one case was relabelled after every examined judge
read it against the key.
## Provenance
Set version `2.3`. Shipped inside the package as
`journeyman/calibration/v2_real.json` — `pip install journeyman-bench`.
Archived releases carry a DOI: [10.5281/zenodo.22085820](https://doi.org/10.5281/zenodo.22085820).
```
journeyman qualify --judge <endpoint> --judge-model <model> --repeats 3
```
## Looking for an arXiv endorser (cs.AI / cs.LG)
The write-up describing this set is in progress. arXiv requires an endorsement
for a first submission in a category and we have no institutional affiliation
to bypass it. If you publish in this area and think the work is worth
endorsing, please open an issue on
[the repository](https://github.com/codechu/journeyman/issues) — and if you
read it and think it is not, that is useful to hear too.
Generated from the set itself by `tools/hf_sync.py`.