rrivera1849/style-aware-paraphraser-outputs

Style-Aware Paraphraser Outputs

⚠️ Intended for evaluating machine-text detectors, not for training them. Each row is a (human, machine, adversarial-paraphrase) triple drawn from a specific evaluation slice of three domains. The author bank that produced these outputs overlaps with the rows here, so training a detector against this data would not generalize. Use it to score a detector, not to fit one.

Final outputs of the style-aware paraphraser from Attacks on Machine-Text Detectors Retain Stylistic Fingerprints (ICML 2026) across three domains, exposed as three named splits. Each row contains a human-written reference, the corresponding machine generation, and our system's final paraphrase after 3 iterations of refinement.

from datasets import load_dataset
ds = load_dataset("rrivera1849/style-aware-paraphraser-outputs")
# ds["reddit"], ds["amazon"], ds["blogs"]

These are the outputs that drive the headline result in the paper: per-sample detection AUROC drops well below baseline at $N{=}1$ across all three domains, yet multi-document detection ($N{\ge}25$) still recovers the signal because the target author's style persists across our paraphrases.

Schema

Each {domain}.jsonl file has three string fields per row:

Field	Description
human_text	Original human-written post (Reddit comment / Amazon review / blog excerpt).
machine_text	Mistral-7B-Instruct-v0.3 generation conditioned on the target author's style exemplars. This is the "machine" side of the detection task.
adversarial_paraphrase	Our system's output: 3 iterations of style-aware refinement applied to a Mistral-7B paraphrase of machine_text, top-1 by SBERT similarity to a target-author exemplar set.

Files

File	Rows	Domain
reddit.jsonl	12,000	Reddit comments (Million User Dataset)
amazon.jsonl	12,000	Amazon product reviews
blogs.jsonl	7,000	Blog Authorship Corpus

How these were produced

For each row we (a) prompt Mistral-7B-Instruct-v0.3 to write a comment / review / post in a target author's style (machine_text); (b) paraphrase machine_text five times with Mistral-7B; (c) run our DPO-tuned paraphraser for three iterations of refinement (10 candidates per iteration, top-5 by SBERT against the target author's exemplars); (d) keep the top-1 pick as our_text.

See §3 and §F of the paper for full details, and the model repo rrivera1849/style-aware-paraphraser-mistral7b for the checkpoint that produced these outputs.

Intended use

• Benchmarking machine-text detectors against a known-difficult paraphrase attack. Compute detector scores on machine_text vs. adversarial_paraphrase; the gap quantifies attack effectiveness at $N{=}1$.
• Studying stylistic-fingerprint detection. Paired (human_text, adversarial_paraphrase) rows let downstream work probe what stylistic features survive aggressive style-aware paraphrasing.

Citation

@inproceedings{rivera-soto-etal-2026-attacks,
  title     = {Attacks on Machine-Text Detectors Retain Stylistic Fingerprints},
  author    = {Rivera Soto, Rafael A. and Chen, Barry and Andrews, Nicholas},
  booktitle = {Proceedings of the International Conference on Machine Learning},
  year      = {2026},
  url       = {https://arxiv.org/abs/2505.14608},
}

License

MIT. The text was generated using mistralai/Mistral-7B-Instruct-v0.3 (Apache 2.0) and our DPO-tuned derivative; consult those model cards for their respective terms.