---
language:
- en
license: other
license_name: mixed-see-source
license_link: https://github.com/uspraveen/Jevify/blob/main/docs/DATASETS.md
task_categories:
- text-classification
tags:
- decision-model
- calibration
- jev
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
  - split: validation
    path: data/validation-*
---
# jevbench

A flat, loadable sample of [Praveenrajus/jev-bench](https://huggingface.co/datasets/Praveenrajus/jev-bench)
(revision `c37b0f6`): 200 rows per data config and split, from the 22 data configs (model-prediction dumps
excluded), merged into one table with the config in a `task` column.

Each row is a decision: a `state`, a typed `question` (JSON: `type` is `choice`, `noul` yes/no, or `score`
on an ordinal scale, with `instructions` and `criteria`), the gold `label`, and `soft_label`, the human label
distribution where annotator votes exist (chaosnli, go_emotions, civil_comments, measuring_hate_speech; empty
otherwise). Every field is a string: the source configs disagree on `soft_label`'s type, so they cannot be
loaded as one dataset directly.

Licenses are those of the underlying datasets; see the source dataset's
[manifest](https://github.com/uspraveen/Jevify/blob/main/docs/DATASETS.md). Rows are sampled with a fixed
seed by `build_jevbench.py` in this repo:
`python build_jevbench.py --out DIR [--push tasksource/jevbench]` rebuilds it exactly.