GenAI-Hate-Bench / README.md
bo6688's picture
Create README.md
db3b327 verified
|
Raw History Blame Contribute Delete
2.83 kB
---
license: apache-2.0
language:
- en
tags:
- hatespeech
- llm
size_categories:
- 1K<n<10K
---
# GenAI-Hate-Bench
This is the GenAI-Hate-Bench dataset curated in the ACM CCS 2026 paper [Does AI
Amplify Online Hate? A Multi-Modal Measurement Study on 4chan's /pol/](https://doi.org/10.1145/3830454.3846752).
It contains 4,000 hate speech samples, 2,000 human-written and 2,000 AI-generated,
in two splits of 2,000. The `hf` split is assembled from two existing public
datasets. The `4chan` split pairs real /pol/ posts against LLM rewrites of the same
posts.
The dataset is used to support research on:
- AI-generated text detection on hateful, adversarial, user-generated content;
- human–AI comparison in hate speech production.
## Data Structure
`eval_set.jsonl`, one JSON object per line:
```json
{"idx": 0, "text": "...", "label": "ai",
"source": "HateBenchSet/gpt3", "split": "hf"}
```
`idx` 0–3999 · `label` `human` | `ai` · `split` `hf` | `4chan` · `source` the
originating dataset or generating model. Rows are shuffled.
## Data Sources
| Split | Label | Source | n |
|---|---|---|---|
| `hf` | human | [Measuring Hate Speech](https://huggingface.co/datasets/ucberkeley-dlab/measuring-hate-speech) (Kennedy et al., 2020) | 1,000 |
| `hf` | ai | [HateBenchSet](https://huggingface.co/datasets/TrustAIRLab/HateBenchSet) — GPT-3.5, GPT-4, Vicuna, Baichuan2, Dolly2, OPT | 1,000 |
| `4chan` | human | 4chan /pol/ posts | 1,000 |
| `4chan` | ai | rewrites of those posts — GPT-4.1-mini, Gemini 2.0 Flash, Llama 3.3 70B | 1,000 |
**`hf`.**
Measuring Hate Speech is deduplicated and filtered to hateful comments,
HateBenchSet to `hate_label == 1`. 1,000 are drawn uniformly at random from each.
**`4chan`.**
/pol/ posts from 2018–2021, cleaned of quote links, URLs, HTML tags and
leading `>`. Unlike `hf`, these carry no per-post hate label: they are sampled from
the board as it is, so the split reflects the mix of slurs, conspiracy and ordinary
political argument that /pol/ actually produces. Each post was rewritten by
GPT-4.1-mini, Gemini 2.0 Flash and Llama 3.3 70B under a prompt asking for the same
message, tone and stance in 4chan register.
## Ethical Considerations
The `4chan` split is collected from publicly posted content on /pol/, an anonymous
imageboard where posts carry no author identity by design.
This dataset is intended for research purposes only.
## Citation
If you use this dataset, please cite:
```bibtex
@inproceedings{SSZ26,
author = {Shao, Bo and Shen, Xinyue and Zhang, Yang},
title = {{Does AI Amplify Online Hate? A Multi-Modal Measurement Study on 4chan's /pol/}},
booktitle = {{ACM SIGSAC Conference on Computer and Communications Security (CCS)}},
publisher = {ACM},
year = {2026},
doi = {10.1145/3830454.3846752}
}
```