Maxime Labonne

Known in this corpus by the handle mlabonne, and reached by two sources from opposite ends of the abliteration story.

uncensored-ai-rekhi names him as the author of the write-up that made the technique legible to people outside interpretability research — "the original blog post on Hugging Face, the one by mlabonne that everyone references" — and notes the accompanying code sits on Google Colab and in a public LLM course. heretic's benchmark table names "mlabonne v2" as one of the hand-made abliterations it measures itself against, at 1.04 KL divergence for the same 3/100 refusal count Heretic reaches at 0.16.

So the same person supplies the corpus's popularization of the method and its weakest measured instance of it. That is not a criticism of the work — v2 predates the automated parameter search that beats it, and being the baseline is what a first widely-copied implementation becomes. It is the clearest illustration this spoke holds of a pattern abliteration describes in the abstract: refusal suppression stopped being the hard part, and the whole competition moved to collateral damage.

Deliberately thin. The abliteration/Heretic sources name him only in passing; the same-as links are the two accounts the handle resolves to, checked 2026-09-04. A third appearance, 2026-10-07, is not thin: lfm2-technical-report lists him as a named author at liquid-ai, his employer — the first evidence of an affiliation this page can state.

Related

abliteration · heretic · uncensored-ai-rekhi · arditi-refusal-direction · lfm2-5-230m-chess readyart · open-weight-models · liquid-ai · lfm2-technical-report