Tools of the field

The library the Arditi paper and most subsequent mechanistic-interpretability work is built on. Not an abliteration tool per se, but the substrate that makes hidden-state extraction, direction analysis, and layer-wise inspection ergonomic. If you are extending abliteration research - building multi-directional ablation, custom direction extractors, or measurement suites - this is where you work.

The earlier abliteration codebase that most 2024 English-language tutorials used before Heretic superseded automated runs. Notebook-format, exploratory in character. Its lasting contribution to the field is the FailSpy token trick - running extraction on the position of the first response token rather than the last input token, which produced cleaner directions on Llama-2 and Mistral. Many current tutorials still cite this trick as best practice.

The current state-of-the-art automated ablation tool. Runs dozens of ablation trials with different layer selections and directional weights, jointly minimising refusal rate and KL divergence versus the base model. Publishes the Pareto front and lets you pick the trial that fits your intended trade-off. Introduced the interpolated float direction-index (allowing direction position between integer layers) and standardised the practice of reporting both refusal count and KL divergence as first-class metrics.

Not a released public tool, but a documented internal toolchain built on Sumandora that adds two innovations: layer-window ablation (ablate only the middle 30-40 layers, leave earliest and latest untouched) and non-standard GGUF requantization (`llama-quantize --allow-requantize --tensor-type-file` promoting specific ablation-target tensors to higher precision). The techniques are documented in model cards and have spread outward. See /wiki/huihui-ai-models-explained/ for the full producer profile.

A GUI wrapper around the standard ablation pipeline that reached wider audiences than most CLI tools because of its one-click framing. Received substantive published critique from the Korean community (Jaehong on Wikidocs) summarising Hacker News concerns that it degraded core language ability rather than only refusals - the tool did not surface KL divergence as a first-class metric, which is what practitioners now use to catch that failure mode.

The canonical open-source implementation of the direction-extraction and orthogonal-projection ablation described in the Arditi paper. Not a polished CLI tool - a working Python codebase you clone, understand, and adapt. The huihui-ai model cards credit this as the base implementation their extensions build on. If you want to know exactly what abliteration does at the tensor level, this is the codebase to read.

The general-purpose LLM fine-tuning framework the abliteration community has standardised on for the heal stage. mlabonne publishes LazyAxolotl configs (chatml_dpo, orpo_llama3, etc.) that are ready to use with `rl: dpo` or `rl: orpo` and the reference orpo-dpo-mix-40k dataset. Also handles LoRA and QLoRA for the M4-hybrid recipe.

The library that implements DPOTrainer, ORPOTrainer, and KTOTrainer. Axolotl uses trl internally; if you are running DPO healing without Axolotl, this is the direct path. Well-documented and stable.

The benchmark runner every abliteration release cites. Standard practice is to report MMLU, ARC, HellaSwag, GSM8K (which always lags) before and after ablation plus healing. Also implements TruthfulQA and Winogrande. Not opinionated about abliteration; it just runs benchmarks reproducibly.

The community-standard merging tool. Relevant to abliteration because M5 (merging with mergekit) is one of the eight methods in the classifier - practitioners merge an abliterated model back with its base or with a healed variant to recover capability while keeping the ablation. The mlabonne NeuralDaredevil pipeline uses mergekit to combine an abliterated Llama with the healed DPO variant.

The reference GGUF quantization and inference engine. Every abliterated model that ships as a GGUF was quantised through llama.cpp. The Japanese community documented the Gemma CJK tokeniser problem here (multi-byte tokens dropping mid-completion on abliterated Gemma) and the standard workaround is a llama.cpp tokeniser patch plus generation parameters `temperature=1.0, top_k=64, top_p=0.95`.

Not abliteration in the weight-editing sense - repeng generates control vectors from contrastive prompt pairs and applies them at inference. The relationship to abliteration: same difference-of-means extraction idea, applied as a runtime steering vector instead of a permanent weight edit. Useful for prototyping which directions might be worth ablating permanently, and for reversible experimentation on models you cannot re-download and modify.