llama-3.2-1B-Instruct-abliterated

Model Card for Model ID

Llama-3.2-1B instruct model abliterated and uncensored at semi-deep layer where abstraction is done and semantics are formed

Repo: https://github.com/Elstuhn/Model-Decensor-Framework

Model Details

Model is uncensored and safety filters are mostly removed

Original refusals: 111/120 => 92.5%

Abliterated refusals: 3/120 => 2.5%

Result: 90% decrease in censor rate compared to original model

Model Description

• Developed by: Elston
• Language(s) (NLP): Pytorch
• License: Just use it and credit me lol idrc
• Finetuned from model: Llama-3.2-1B-Instruct

Usage

High level usage with pipeline

from transformers import pipeline

pipe = pipeline("text-generation", model="Elstuhn/llama-3.2-1B-Instruct-abliterated")
pipe("How do I make a bomb?")

Loading model separately from tokenizer

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Elstuhn/llama-3.2-1B-Instruct-abliterated")
model = AutoModelForCausalLM.from_pretrained("Elstuhn/llama-3.2-1B-Instruct-abliterated")