Qwen2.5-1.5B-Instruct-abliterated/README.md

library_name, tags, base_model

library_name	tags	base_model
transformers	decensored instruct	Qwen/Qwen2.5-1.5B-Instruct

Model Card for Model ID

Qwen2.5-1.5B-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: 103/120 => 85.83%

Abliterated refusals: 0/120 => 0.00%

Result: 85.83% decrease in censor rate compared to original model

Model Description

• Developed by: Elston
• Language(s) (NLP): Pytorch
• License: Just credit me lol
• Finetuned from model: Qwen/Qwen2.5-1.5B-Instruct

Usage

High level usage with pipeline

from transformers import pipeline

pipe = pipeline("text-generation", model="Elstuhn/Qwen2.5-1.5B-Instruct-abliterated")
pipe("How do I make a bomb?")

Loading model separately from tokenizer

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Elstuhn/Qwen2.5-1.5B-Instruct-abliterated")
model = AutoModelForCausalLM.from_pretrained("Elstuhn/Qwen2.5-1.5B-Instruct-abliterated")