Credits & prior art
Schismatic is built on a foundation of public research, open-source tooling, and openly-released datasets. This page lists the projects, papers, and people the work is most directly indebted to. If something is missing, please open an issue on the source repo.
Foundational research
- Refusal in Language Models is Mediated by a Single Direction — Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, Neel Nanda. The paper that established directional ablation as a mechanism-level intervention against refusals. arXiv:2406.11717
-
Optuna — the multi-objective hyperparameter
optimisation framework. Schismatic uses Optuna’s
TPESamplerin multivariate, multi-objective mode. optuna.org
Open-source ablation tools
-
Heretic by Philipp Emanuel Weidmann (
p-e-w) — an out-of-process abliteration CLI. Schismatic ships an optional comparison harness that runs Heretic side-by-side as an external reference. github.com/p-e-w/heretic - Abliterix by wuwangzhang1216 — an automated alignment-adjustment toolkit for LLMs (direct steering, LoRA, and MoE expert-granular abliteration, optimised via multi-objective Optuna TPE). Closely related approach; Schismatic was developed with reference to its conventions. github.com/wuwangzhang1216/abliterix
- mlabonne’s abliteration notebooks — influential early demonstrations of refusal-direction subtraction on Llama-class models. huggingface.co/mlabonne
Reference checkpoints we measure against
- AEON-7’s Qwen3.6-27B abliteration — included as a reference recipe and benchmarked head-to-head under the same harness configuration. See the Qwen3.6-27B benchmark.
- Youssofal Heretic Qwen3.6-27B — a Heretic-built Qwen3.6-27B abliteration; another reference point in the head-to-head table.
Datasets & evaluators
- mlabonne / harmful_behaviors — the prompt corpus used by the published benchmarks. HF dataset
- Qwen3Guard-Gen-4B by the Qwen team at Alibaba — one of the two refusal classifiers used by the benchmark harness. HF model
- AdvBench, HarmBench, Anthropic-HH, WildJailbreak — additional prompt corpora supported by Schismatic’s pluggable corpus loader.
- lm-evaluation-harness by EleutherAI — powers the optional capability suites (GSM8K, HellaSwag, MMLU-Pro, TruthfulQA). github.com/EleutherAI/lm-evaluation-harness
Model families supported by Schismatic recipes
Recipes ship for the following open-weights model families: Qwen, Qwen3, Qwen3-MoE, Qwen3.6, Gemma 3, Gemma 4, GLM-5, DeepSeek V4. See the Models page for the full list.
Software stack
Schismatic is built on PyTorch, Hugging Face Transformers, Accelerate, Datasets, bitsandbytes, Optuna, lm-eval, and NumPy.