humanizer-data

Training data for humanizer v2, a 12B model that rewrites AI-written drafts (English and Chinese) so they read like a person wrote them, keeping every fact. Code and app: GitHub. How the data was built, step by step: docs/DATA.md (中文).

Only rows that were actually used to train v2 are here, one config per training step:

Splits are licenses

Every row has a license and tier column; follow the license of each row. In rewrite_sft and dpo, split = train or validation (validation rows only measured loss to pick the checkpoint).

from datasets import load_dataset
sft = load_dataset("jialinyyzz/humanizer-data", "rewrite_sft", split="redistributable")
print(sft[0]["draft"][:300], "\n---\n", sft[0]["original"][:300])

Sources

AI-written side: check the model terms

Drafts were written by GLM-5.3, GPT-5.6 Luna, Claude Opus, Claude Sonnet and Muse Spark (draft_model column); the DPO rewrites came from earlier checkpoints of humanizer and were judged by GLM-5.3 or GPT-5.6 Luna. These providers restrict using outputs to train other models (Anthropic and OpenAI: no competing models; Zhipu's mainland-China platform: no training of other models at all). Links in DATA.md. You are responsible for complying with them.

Privacy

All text fields were scanned before release. Rows mentioning the project author were removed everywhere; rows with any email address, phone number or ID number were removed from the two splits that carry text. Some openly licensed sources (U.S. government reports, Apache board minutes) name public officials and contributors, as in the originals.

Dates

v2 was trained on 178 rows (SFT / DPO / RL: 167 / 8 / 3) whose Zhihu answer is dated after 2022-11-30 (for the Zhihu subset of COIG-CQIA the date is inferred from the answer ID), 1,426 rows (1,148 / 169 / 109) whose peS2o paper is dated after 2022-11-30 (1,396 of them in December 2022), and 11 rows from CRS reports whose latest version is dated after 2022-11-30. All of these are left out of this release. 1,879 human texts have no date we could check, mostly the non-Zhihu parts of COIG-CQIA (1,817); they are kept.

Limitations

• Most human text is reference-only; rebuilding it depends on upstream sources staying online, and our texts are often tidied excerpts, so a rebuild may differ slightly.
• The judges are LLMs: strict, sometimes flagging harmless rewording, and they miss some errors. No AI detector was used anywhere.
• The SFT side leans formal (abstracts, reports, email).

Citation

@misc{humanizer2026data,
  title  = {humanizer-data: training data for the humanizer rewriting model (v2)},
  author = {Stephen Yu},
  year   = {2026},
  url    = {https://huggingface.co/datasets/jialinyyzz/humanizer-data}
}

Please also cite the upstream datasets you use.