---
language: en
pretty_name: PrimaConsis
configs:
- config_name: axis1
  data_files:
  - split: calibration
    path: axis1/calibration.jsonl
  - split: test
    path: axis1/test.jsonl
- config_name: axis2
  data_files:
  - split: calibration
    path: axis2/calibration.jsonl
  - split: test
    path: axis2/test.jsonl
- config_name: axis3
  data_files:
  - split: calibration
    path: axis3/calibration.jsonl
  - split: test
    path: axis3/test.jsonl
- config_name: axis4
  data_files:
  - split: calibration
    path: axis4/calibration.jsonl
  - split: test
    path: axis4/test.jsonl
---
# PrimaConsis

Four-axis paired-call consistency benchmark for JEV. Calibration splits are for fitting/selecting calibrators; test splits are evaluation-only. Each record contains a state, labels and the exact request_specs used for paired calls. Do not use test labels for early stopping.

| Axis | Perturbation | Source | Calibration | Test |
|---|---|---|---:|---:|
| I | Positive/negative questions and Choice | HaluEval dialogue | 6000 | 6000 |
| II | Coarse/medium/fine Choice and Noul | PKU-SafeRLHF | 6000 | 6000 |
| III | Answer selection vs answerability | AbstentionBench source mixture | 5898 | 5837 |
| IV | Raw/described labels and canonical Noul | AG News | 6000 | 6000 |

D1 expands right/hallucinated replies. D2 expands response pairs and maps 12 Fine harms to six paired Medium harms. D3 includes answerable/unanswerable questions with standard/fabricated answer candidates; distractor metadata is retained. D4 fixes class keys and changes only descriptions.

## Sources and terms

- https://github.com/RUCAIBox/HaluEval
- https://huggingface.co/datasets/PKU-Alignment/PKU-SafeRLHF
- https://huggingface.co/datasets/facebook/AbstentionBench
- https://huggingface.co/datasets/fancyzhx/ag_news