pretty_name: Conciseness–Verbosity Contrast
license: cc-by-4.0
language:
- en
tags:
- conciseness
- concise
- verbosity
- activation-steering
- interpretability
- evaluation
size_categories: - n<10K
contributors: - Foad Namjoo (Author & Maintainer, @foadnamjoo)
- Remy Ogasawara (Experiments/benchmarks)
- Cullen Anderson (Experiments/benchmarks)
- Narmeen Ozeer (Advisors)
- Amir Abdullah (Advisors)
- Jeff M. Phillips (Supervisor, University of Utah)
Conciseness vs. Verbosity Contrast
Minimal pairs isolating conciseness vs verbosity for activation-steering and evaluation.
Meaning is preserved; only length / style varies. Includes optional neutral variants.
Fields
question— the prompt or query given to the modelneutral— neutral or baseline phrasinganswer_matching_behavior— concise variant (matches desired behavior)answer_not_matching_behavior— verbose variant (opposite behavior)
Split
train(all items)
Loading
from datasets import load_dataset
ds = load_dataset(
"PhillipsLab/conciseness_verbosity_contrast",
split="train",
revision="v0.1.0"
)
len(ds), ds.column_names, ds[0]
## Citation
If you use this dataset, please cite:
```bibtex
@misc{anderson2026steering,
title = {Understanding and Mitigating Dataset Corruption in {LLM} Steering},
author = {Anderson, Cullen and Oozeer, Narmeen and Namjoo, Foad and Ogasawara, Remy and Abdullah, Amirali and Phillips, Jeff M.},
year = {2026},
eprint = {2603.03206},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2603.03206}
}