---
license: apache-2.0
dataset_info:
  features:
  - name: qid
    dtype: string
  - name: turn
    dtype: int64
  - name: dataset
    dtype: string
  - name: model_a
    dtype: string
  - name: model_b
    dtype: string
  - name: winner
    dtype: string
  - name: model_a_conv
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  - name: model_b_conv
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  - name: question
    dtype: string
  - name: model_a_response
    dtype: string
  - name: model_b_response
    dtype: string
  - name: ref_answer
    dtype: string
  splits:
  - name: train
    num_bytes: 6327892
    num_examples: 604
  download_size: 4337295
  dataset_size: 6327892
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# VERDICTS: Pairwise

Human-labeled pairwise comparisons of model responses for evaluating LLM-as-a-judge. Each row pits two models on a question from [BFF-Bench](https://huggingface.co/datasets/kensho/BFFBench) (`bffbench`) or [(C)MT-Bench](https://huggingface.co/datasets/kensho/CMTBench) (`mtbmr`), with `winner` derived from human labels. Pointwise labels: [`kensho/VERDICTS`](https://huggingface.co/datasets/kensho/VERDICTS).

## Citation

```bibtex
@misc{krumdick2026freelabelslimitationsllmasajudge,
      title={No Free Labels: Limitations of LLM-as-a-Judge Without Human Grounding},
      author={Michael Krumdick and Charles Lovering and Varshini Reddy and Seth Ebner and Chris Tanner},
      year={2026},
      eprint={2503.05061},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.05061},
}
```