---
license: apache-2.0
dataset_info:
  features:
  - name: qid
    dtype: string
  - name: cid
    dtype: string
  - name: turn
    dtype: int64
  - name: model
    dtype: string
  - name: label
    dtype: string
  - name: time
    dtype: float64
  - name: dataset
    dtype: string
  - name: conv
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  splits:
  - name: train
    num_bytes: 12193053
    num_examples: 3600
  download_size: 11980467
  dataset_size: 12193053
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# VERDICTS

Human labels of model-response correctness for evaluating LLM-as-a-judge. Each row is a model's response to a question from [BFF-Bench](https://huggingface.co/datasets/kensho/BFFBench) (`bffbench`) or [(C)MT-Bench](https://huggingface.co/datasets/kensho/CMTBench) (`mtbmr`), with a human `label`. Pairwise comparisons: [`kensho/VERDICTSPairwise`](https://huggingface.co/datasets/kensho/VERDICTSPairwise).

## 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},
}
```