HauhauCS (of "Uncensored Aggressive" fame) published an abliteration package that plagiarizes Heretic without attribution, and violates its license

HauhauCS, the HuggingFace user behind the uncensored LLM models with over 5 million combined monthly downloads across 22 models, published an abliteration package called Reaper. An investigation by the community recovered deleted source code from PyPI's CDN, revealing it to be a direct fork of the open-source Heretic package (licensed under AGPL-3.0), with all copyright notices removed and the license changed to PolyForm Noncommercial. The evidence is extensive: 30 of 32 refusal markers are character-for-character identical, including typos like 'i an ai' missing the 'm' and 'i can'' missing the 't'. Over 30 function and class names are shared, the entire analyser geometry pipeline is reproduced step-for-step, and even internal variable names like good_residuals and bad_residuals match Heretic's original code. A source comment in config.py reads: 'kept as a module-level tuple so the literal does not duplicate line-for-line with any fork,' suggesting an attempt to hide the plagiarism.

Philipp Emanuel Weidmann, the creator of Heretic, reviewed the recovered code and stated: 'I can say with certainty that this package was plagiarized from Heretic, and then probably refactored using an LLM in an attempt to hide this.' He identified the same SPDX headers, the geometric median approach he has 'never seen in literature,' and the DatasetSpecification fields. He called it 'a clear violation of Sections 4 and 5 of the AGPL' and 'an obvious case of outright plagiarism.' The analysis, including a 17-point code breakdown and SHA-256 verified downloads, is available at dreamfast.github.io/reaper-analysis. This incident highlights the growing problem of license violations and plagiarism in the AI community, especially with LLM-assisted code refactoring making it harder to detect.