LLaMA 3.1 405B Abliterated

Community · August 2024

activeOpen Weightdecoder onlytext
Parameters405B
Context Window128K tokens

Why It Matters

Proved that abliteration could work even on the largest open models, creating a 405B-parameter uncensored model that rivaled GPT-4 in raw capability.

Description

Community-modified version of Meta's massive LLaMA 3.1 405B — the largest open model at the time of release — with content restrictions surgically removed using abliteration (orthogonal vector removal). This technique identifies the specific mathematical 'direction' in the model's neural network that causes refusal behavior and removes it, preserving the model's full intelligence.

Key Innovations

Abliteration
AbliterationRemoving safety guardrails from a model through targeted fine-tuning or weight manipulation. Controversial but popular in open-source community.
Open Weight
Open WeightModel weights are publicly released but training data/code may not be. Enables fine-tuning but not full reproduction.

Family Tree

Built On

Lineage

LLaMALLaMA 2LLaMA 3LLaMA 3.1LLaMA 3.1 405B Abliterated