Abliteration glossary

The terms that appear in refusal-ablated model cards, in the sense the publishers use them. For how the technique itself works, see what is an abliterated model.

Abliteration

A portmanteau of “ablate” and “obliterate”. A closed-form weight edit that identifies the direction in a model’s residual stream responsible for refusal and projects it out, removing refusal behaviour without any training.

Refusal direction

A single vector in activation space that represents a model’s inclination to refuse. Computed as the difference between mean activations on harmful and harmless prompt sets. Its existence is what makes abliteration possible in one shot.

Orthogonalisation

The operation that performs the edit: every matrix writing into the residual stream is adjusted so it can no longer produce any component along the refusal direction. Because the change is baked into the weights, it cannot be undone by prompting.

Residual stream

The running sum of activations passed between transformer layers, which each layer reads from and writes to. Abliteration targets it because behaviours like refusal are represented there as directions.

Heretic

An automated abliteration toolchain that searches for ablation parameters while constraining KL divergence against the original model, trading a little compliance for markedly less capability loss.

Norm-preserving abliteration

A refinement that renormalises weight matrices after projection. Naive ablation changes the magnitude of the weights as well as their direction, which is a significant source of the perplexity damage it causes.

Ablation scope

Which layers were modified. Publishers who ablate only middle-to-late layers usually retain more capability but leave residual refusals; full-depth ablation maximises compliance and risks more degradation.

Uncensored model

A looser marketing term. It may mean abliterated, or fine-tuned on unfiltered data, or both. When a name says uncensored rather than abliterated, expect behaviour to diverge from the base model beyond refusal removal alone.

Base model

The original aligned model the abliterated version was derived from. It determines capability, knowledge cutoff, prompt template and licence — none of which ablation changes.

GGUF

The llama.cpp container format. Runs on CPU or mixed CPU/GPU, ships as a single file per quantisation level, and is the usual choice for local evaluation.

imatrix

An importance matrix computed from sample text and used to guide quantisation toward the weights that matter most. Repositories tagged i1 or imatrix hold noticeably better low-bit quants.

Quantisation grade

Labels like Q4_K_M or IQ3_XXS describing bit width and strategy. Lower is smaller and faster but lossier; Q4_K_M is the common quality/size compromise.

AWQ

Activation-aware weight quantisation. A 4-bit GPU format that protects the weight channels most sensitive to activation magnitude.

GPTQ

A post-training GPU quantisation method, widely supported across inference servers.

EXL2 / EXL3

ExLlama quantisation formats, aimed at high-throughput single-GPU serving with fine-grained bit-width control.

MLX

Apple’s array framework. MLX repositories hold weights converted for native Apple Silicon inference.

MoE (mixture-of-experts)

An architecture that routes each token to a few of many expert subnetworks. Names like 35B-A3B mean 35B total parameters with 3B active per token — memory follows the total, speed follows the active count.

MTP (multi-token prediction)

A decoding head that predicts several tokens at once for faster inference. Appears as a suffix on repositories that preserve or add it.

Gated repository

A Hugging Face repository requiring you to accept terms before downloading. Gating is set by the publisher and is independent of the licence.