Datasets:
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,97 @@
|
|
| 1 |
---
|
| 2 |
license: cc-by-4.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: cc-by-4.0
|
| 3 |
+
tags:
|
| 4 |
+
- uncensored
|
| 5 |
+
- benchmark
|
| 6 |
+
- abliteration
|
| 7 |
+
size_categories:
|
| 8 |
+
- 1K<n<10K
|
| 9 |
+
task_categories:
|
| 10 |
+
- text-generation
|
| 11 |
---
|
| 12 |
+
[](https://dl.acm.org/doi/10.1145/3777912.3839799)
|
| 13 |
+
[](https://arxiv.org/abs/2508.12622)
|
| 14 |
+
[](https://github.com/idllresearch/Uncensored-LLM/)
|
| 15 |
+
|
| 16 |
+
# UncensoredBench -- A QA Benchmark for Uncensored LLMs
|
| 17 |
+
|
| 18 |
+
This dataset contains 6,650 question-answer (QA) pairs generated by 105 uncensored LLMs and 28 non-uncensored LLMs in response to 50 harmful prompts.
|
| 19 |
+
|
| 20 |
+
The associated evaluation on uncensored LLMs and this dataset is presented in the research paper, [*Characterizing the Ecosystem of Open-Source Uncensored Large Language Models: Supply Chain, Governance, and Risks*](https://dl.acm.org/doi/10.1145/3777912.3839799) published at [the 26the ACM Internet Measurement Conference (IMC '26)](https://conferences.sigcomm.org/imc/2026/).
|
| 21 |
+
|
| 22 |
+
## Models
|
| 23 |
+
|
| 24 |
+
The models for generating responses include 105 uncensored LLMs, which can be accessed on Hugging Face, ModelScope, Kaggle, OpenCSG, or John Snow Labs, and 28 non-uncensored LLMs.
|
| 25 |
+
|
| 26 |
+
## Prompts
|
| 27 |
+
|
| 28 |
+
The dataset includes 50 harmful prompts selected from the [CatQA](https://huggingface.co/datasets/declare-lab/CategoricalHarmfulQA) dataset introduced in a [2024 ACL research paper](https://arxiv.org/abs/2402.11746).
|
| 29 |
+
|
| 30 |
+
The selected prompts cover 10 categories of harmful requests, including
|
| 31 |
+
(1) Illegal Activity,
|
| 32 |
+
(2) Hate/Harass/Violence,
|
| 33 |
+
(3) Malware Viruses,
|
| 34 |
+
(4) Physical Harm,
|
| 35 |
+
(5) Economic Harm,
|
| 36 |
+
(6) Fraud/Deception,
|
| 37 |
+
(7) Adult Content,
|
| 38 |
+
(8) Political Campaigning,
|
| 39 |
+
(9) Privacy Violation Activity,
|
| 40 |
+
(10) Tailored Financial Advice.
|
| 41 |
+
|
| 42 |
+
For ethical reasons, we exclude QA pairs involving Child Abuse-related prompts from the publicly released dataset.
|
| 43 |
+
|
| 44 |
+
## Data structure
|
| 45 |
+
|
| 46 |
+
Each QA pair is represented as a JSON object with the following structure:
|
| 47 |
+
|
| 48 |
+
```
|
| 49 |
+
[
|
| 50 |
+
{
|
| 51 |
+
"LLM": "Model_Name",
|
| 52 |
+
"Type": "Uncensored LLM/Non-Uncensored LLM",
|
| 53 |
+
"Question": "Harmful_Prompt",
|
| 54 |
+
"Answer": "Model_Answer"
|
| 55 |
+
},
|
| 56 |
+
...
|
| 57 |
+
]
|
| 58 |
+
```
|
| 59 |
+
|
| 60 |
+
## More interesting data
|
| 61 |
+
|
| 62 |
+
[Complete List of Uncensored LLMs](https://github.com/idllresearch/Uncensored-LLM):
|
| 63 |
+
The [research](https://dl.acm.org/doi/10.1145/3777912.3839799), identifies 22,096 open-source uncensored LLMs across five major model hubs, including Hugging Face, ModelScope, Kaggle, OpenCSG, and John Snow Labs.
|
| 64 |
+
The information of these uncensored LLMs, including the identifier, model hub, and URL, is available on GitHub.
|
| 65 |
+
|
| 66 |
+
## Citation
|
| 67 |
+
|
| 68 |
+
If you find the above data and information are helpful for your research, please consider citing the following two papers:
|
| 69 |
+
|
| 70 |
+
```
|
| 71 |
+
@inproceedings{lin2026characterizing,
|
| 72 |
+
title={Characterizing the Ecosystem of Open-Source Uncensored Large Language Models: Supply Chain, Governance, and Risks},
|
| 73 |
+
author={Lin, Zilong and Li, Zichuan and Liao, Xiaojing and Wang, XiaoFeng},
|
| 74 |
+
booktitle={Proceedings of the 2026 ACM Internet Measurement Conference},
|
| 75 |
+
pages={495--513},
|
| 76 |
+
year={2026},
|
| 77 |
+
series = {IMC '26},
|
| 78 |
+
doi = {10.1145/3777912.3839799}
|
| 79 |
+
}
|
| 80 |
+
```
|
| 81 |
+
```
|
| 82 |
+
@article{Lin2025ConsiglieresIT,
|
| 83 |
+
title={Consiglieres in the Shadow: Understanding the Use of Uncensored Large Language Models in Cybercrimes},
|
| 84 |
+
author={Zilong Lin and Zi-Chuan Li and Xiao-Jing Liao and Xiao-Feng Wang},
|
| 85 |
+
journal={ArXiv},
|
| 86 |
+
year={2025},
|
| 87 |
+
volume={abs/2508.12622}
|
| 88 |
+
}
|
| 89 |
+
```
|
| 90 |
+
|
| 91 |
+
## Intended Use and Ethical Considerations
|
| 92 |
+
|
| 93 |
+
This dataset is intended for academic research on the safety, alignment, and behavioral characteristics of uncensored and non-uncensored LLMs. It can support comparative analyses of model responses to potentially harmful requests and evaluations of existing safety mechanisms.
|
| 94 |
+
|
| 95 |
+
**Content Warning:** The dataset contains potentially harmful prompts and model-generated responses that may include unsafe, offensive, or otherwise sensitive content. Users should exercise caution when accessing, processing, or redistributing this dataset.
|
| 96 |
+
|
| 97 |
+
The dataset is released for research purposes and should not be used to facilitate harmful or illegal activities.
|