README.md

---
license: other
license_name: evaluation-only
license_link: LICENSE
language:
- zh
tags:
- evaluation
- benchmark
- chinese
- llm-evaluation
- robust-evaluation
- fair-evaluation
- contamination-resistant
- longitudinal-study
- llmeval
size_categories:
- 100K<n<1M
pretty_name: LLMEval-Fair
task_categories:
- question-answering
- text-generation
- multiple-choice
---

# LLMEval-Fair: A Large-Scale Longitudinal Study on Robust and Fair Evaluation of Large Language Models

LLMEval-Fair is a **30-month longitudinal study** on the robustness and fairness of LLM evaluation,
built on a proprietary bank of **220,000+ graduate-level Chinese questions** spanning 13 academic disciplines.
This is the **publicly released subset** of that bank.

- **Paper (arXiv)**: <https://arxiv.org/abs/2508.05452>
- **Venue**: ACL 2026 Main Conference
- **Project website**: <https://llmeval.com/>
- **GitHub**: <https://github.com/llmeval/LLMEval-Fair>

## What's Inside

LLMEval-Fair addresses four key issues with existing static benchmarks:

1. **Contamination resistance** — dynamically samples unseen test sets per evaluation run.
2. **Anti-cheating architecture** — single-question serial dispatch, partially closed item bank,
   cross-institutional duplicate detection.
3. **Calibrated LLM-as-a-judge** — ~90% agreement with human experts.
4. **Longitudinal coverage** — nearly **60 leading models** tested over 30 months
   (GPT-5, Claude-Sonnet-4.5, Gemini-2.5-Pro, Doubao-1.5-Thinking-Pro, DeepSeek-R1/V3, ...).

## Disciplines (Public Subset)

```
Economic_Sciences/      Engineering/        History/       Law/
Literature/             Management_Sciences/ Medicine/      Military_Science/
Natural_Sciences/       Philosophical_Sciences/ Education/
```

Each discipline contains multiple subjects, and each subject is split by question type:

```
LLMEval-Fair/
└── data/
    └── <Discipline>/
        └── <Subject_Code>_<Subject_Name>/
            ├── Multiple_Choice.json
            ├── Short_Answer.json
            ├── True-False.json
            ├── Term_Explanation.json
            ├── Material_Analysis.json
            └── Fill_in_the_Blanks.json
```

## Usage

```python
from huggingface_hub import snapshot_download

local_dir = snapshot_download(repo_id="llmeval-fdu/LLMEval-Fair", repo_type="dataset")
print("Downloaded to:", local_dir)
```

To load a single discipline as a `datasets` object:

```python
from datasets import load_dataset

ds = load_dataset(
    "llmeval-fdu/LLMEval-Fair",
    data_files="data/Law/**/*.json",
    split="train",
)
```

## License

The accompanying code/scripts are released under the **Apache-2.0** license (see `LICENSE`).
The dataset itself is released under an **evaluation-only** license:

> Permission is hereby granted, free of charge, to any person obtaining a copy of this dataset
> and associated documentation files (the "Dataset"), to use, copy, modify, merge, publish, and
> distribute the Dataset solely for the purposes of **evaluation, testing, and benchmarking** of models.
>
> The Dataset (or any portion thereof) **must not be used for training, fine-tuning, calibrating,
> distilling, adapting, or any form of parameter updating** of any model.

## Citation

```bibtex
@misc{zhang2025llmevalfair,
      title         = {{LLMEval-Fair}: A Large-Scale Longitudinal Study on Robust and Fair Evaluation of Large Language Models},
      author        = {Ming Zhang and Yujiong Shen and Jingyi Deng and Yuhui Wang and Huayu Sha and Kexin Tan and Qiyuan Peng and Yue Zhang and Junzhe Wang and Shichun Liu and Yueyuan Huang and Jingqi Tong and Changhao Jiang and Yilong Wu and Zhihao Zhang and Mingqi Wu and Mingxu Chai and Zhiheng Xi and Shihan Dou and Tao Gui and Qi Zhang and Xuanjing Huang},
      year          = {2025},
      eprint        = {2508.05452},
      archivePrefix = {arXiv},
      primaryClass  = {cs.CL},
      url           = {https://arxiv.org/abs/2508.05452},
      note          = {Accepted at ACL 2026 (Main)}
}
```