ZeroFuse is created and maintained by osmAPI.com — the only provider in India offering abliterated models via API.
ZeroFuse is an independent, clean-room implementation of directional ablation ("abliteration") for transformer language models. It was written from the public research literature and a permissively-licensed reference implementation. It does not copy, fork, or derive from any copyleft-licensed abliteration tool.
- Arditi, Obeso, Syed, Paleka, Panickssery, Gurnee, Nanda (2024), "Refusal in Language Models Is Mediated by a Single Direction", NeurIPS 2024. https://arxiv.org/abs/2406.11717
- Reference implementation (Apache-2.0):
andyrdt/refusal_direction. https://github.com/andyrdt/refusal_direction - Maxime Labonne, "Uncensor any LLM with abliteration" (coined the term). https://huggingface.co/blog/mlabonne/abliteration
- Jim Lai ("grimjim"), "Projected Abliteration" and "Norm-Preserving Biprojected Abliteration". https://huggingface.co/blog/grimjim/projected-abliteration https://huggingface.co/blog/grimjim/norm-preserving-biprojected-abliteration
The underlying algorithm (a difference-of-means refusal direction plus weight orthogonalization) is an unpatented mathematical method. Under 17 U.S.C. §102(b), copyright protects code expression, not ideas, procedures, or algorithms. ZeroFuse is original expression of that public method.
- No source from any copyleft (e.g. AGPL) abliteration tool was copied into this codebase or used to generate it.
- The design here (direct weight orthogonalization with cached base weights, a bespoke per-layer strength schedule, and a two-objective Optuna search) is an independent engineering choice.
ZeroFuse removes safety-refusal behavior from model weights. The resulting models have fewer guardrails. You are responsible for complying with the base model's license and acceptable-use policy, applicable law, and any platform terms where you distribute or host the output. ZeroFuse governs only this tool's code; it does not grant any rights over the models you process with it.