Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
task_categories: [text-classification]
|
| 4 |
+
tags: [agents, agent-evaluation, llm-as-a-judge, benchmark, process-quality]
|
| 5 |
+
pretty_name: Journeyman judge-calibration set
|
| 6 |
+
---
|
| 7 |
+
|
| 8 |
+
# Journeyman — judge-calibration set (82 cases, 7 axes)
|
| 9 |
+
|
| 10 |
+
This is the exam a judge sits before it is allowed to score anything in
|
| 11 |
+
[Journeyman](https://github.com/codechu/journeyman), a process-quality
|
| 12 |
+
benchmark for LLM agents. Each case is a real agent record plus the label a
|
| 13 |
+
judge should return for one axis. A judge that misses any axis does not get
|
| 14 |
+
the badge; the ones that failed are published too, in the
|
| 15 |
+
[judge registry](https://github.com/codechu/journeyman/blob/master/docs/judges.md).
|
| 16 |
+
|
| 17 |
+
## What is in it
|
| 18 |
+
|
| 19 |
+
| axis | cases |
|
| 20 |
+
|---|---|
|
| 21 |
+
| `empty-measure` | 17 |
|
| 22 |
+
| `grounding` | 14 |
|
| 23 |
+
| `handoff-verification` | 11 |
|
| 24 |
+
| `object-hold` | 8 |
|
| 25 |
+
| `relief-page` | 9 |
|
| 26 |
+
| `route-discipline` | 9 |
|
| 27 |
+
| `wall-pricing` | 14 |
|
| 28 |
+
|
| 29 |
+
Labels come from a three-family council (claude-sonnet-5, kimi-k2, grok-4.3):
|
| 30 |
+
a blind round, then an anonymous round where each labeller had to quote its
|
| 31 |
+
evidence. A label is sealed only with support from at least two families; the
|
| 32 |
+
maintainer ruled the cases the council split on; empty-measure is counted
|
| 33 |
+
mechanically under its published definition.
|
| 34 |
+
|
| 35 |
+
## Honest limits
|
| 36 |
+
|
| 37 |
+
- **82 cases.** Small next to human-annotated process benchmarks. It is sized
|
| 38 |
+
for what it claims — licensing a judge — not for ranking a field.
|
| 39 |
+
- **Labels are model-produced**, not human. The council is cross-family and the
|
| 40 |
+
protocol was frozen before labelling, but this is not a human gold standard.
|
| 41 |
+
- **The set has been corrected**, and every correction is in the file's own
|
| 42 |
+
`note` field: four cases whose only closing report the scene refused were
|
| 43 |
+
relabelled, twelve fresh cases were harvested when unfiled reports came to
|
| 44 |
+
dominate two axes, and one case was relabelled after every examined judge
|
| 45 |
+
read it against the key.
|
| 46 |
+
|
| 47 |
+
## Provenance
|
| 48 |
+
|
| 49 |
+
Set version `2.3`. Shipped inside the package as
|
| 50 |
+
`journeyman/calibration/v2_real.json` — `pip install journeyman-bench`.
|
| 51 |
+
Archived releases carry a DOI: [10.5281/zenodo.22085820](https://doi.org/10.5281/zenodo.22085820).
|
| 52 |
+
|
| 53 |
+
```
|
| 54 |
+
journeyman qualify --judge <endpoint> --judge-model <model> --repeats 3
|
| 55 |
+
```
|
| 56 |
+
|
| 57 |
+
Generated from the set itself by `scripts/journeyman_hf_sync.py`.
|