Abliteration glossary
This abliteration glossary defines the terms you meet when choosing an abliterated model or API. For the method itself, read what abliteration is.
- Abliteration
- Removing a model's refusal direction from its weights, with no retraining. The word blends "ablate" and "obliterate".
- Refusal direction
- A direction in a model's activations that makes it refuse, which Arditi et al. found across open chat models in 2024.
- Orthogonalization
- The linear-algebra step that edits each weight matrix so it can no longer write along the refusal direction.
- Heretic
- An open-source tool that searches for the abliteration settings that remove the most refusals with the least damage.
- Obliterated
- A common misspelling of abliterated.
- Uncensored fine-tune
- A model retrained on data with refusals filtered out, such as Dolphin or Hermes.
- Jailbreak
- A prompt that talks a model out of refusing, with no change to its weights.
- GGUF
- The model file format that llama.cpp, Ollama and LM Studio load.
- Quantization
- Storing weights in fewer bits, so a 4-bit build is under a third the size of the 16-bit original.
- Mixture of experts
- A model that runs only some of its parameters per token, such as 18B of GLM-5.3 Flash's 320B.
- Context window
- How many tokens a model reads in one request, prompt and reply together.
- Zero data retention
- A provider's promise not to store prompts and outputs after a request.
- Blended price
- A price per 1M tokens that weights input and output 3 to 1.
- Anthropic Messages API
- The request format Claude Code uses, at /v1/messages.
- OpenAI Responses API
- The newer OpenAI request format that Codex uses, at /responses.