---
license: apache-2.0
task_categories:
- text-classification
- text-generation
language:
- en
tags:
- ai-generated-text-detection
- adversarial-robustness
- paraphrase-attacks
- stealthrl
- benchmark
pretty_name: StealthRL Benchmark
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.parquet
---

# StealthRL Benchmark

A paired benchmark for evaluating AI-text detectors under adversarial rewriting.
Each row contains an AI-generated source text and a StealthRL rewrite.

Links:

- Paper: https://arxiv.org/abs/2602.08934
- Model: https://huggingface.co/suraj-ranganath/StealthRL
- Demo: https://stealthrl.pages.dev/
- Source dataset: https://huggingface.co/datasets/yaful/MAGE

## Dataset construction

This benchmark uses the same filtered MAGE test-pool AI examples used in the
StealthRL paper. We start from the MAGE test split, apply the paper's evaluation
filtering, and keep the **14,656 AI-generated examples** used for the full
detector evaluation.

For each source example, we include the original AI-generated text and a
StealthRL rewrite generated directly with the released StealthRL model.

## Columns

- `sample_id`: stable row identifier from the filtered MAGE test pool.
- `source`: raw MAGE source field.
- `source_domain`: parsed source/domain family, e.g. `imdb`, `xsum`, `squad`, `yelp`, `cmv`.
- `source_generator`: parsed original generator when available, e.g. `gpt4`, `text-davinci-003`, `gpt-3.5-trubo`.
- `source_task`: parsed generation mode when available, e.g. `continuation`, `specified`, `topical`, `paraphrase`, `direct`.
- `ai_generated_text`: original AI-generated text evaluated in the paper.
- `stealthrl_text`: StealthRL-model rewrite of the AI-generated text.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("suraj-ranganath/StealthRL-Benchmark")["train"]
```

Use `ai_generated_text` to measure clean AI-text detection and `stealthrl_text`
to measure detector robustness to StealthRL adversarial paraphrases.

## Provenance

The source examples come from MAGE (`yaful/MAGE`, Apache-2.0). The subset
corresponds to the filtered full-MAGE evaluation in the StealthRL paper.