A meme tool to perform "digital lobotomy" on any LLM, forcing it to only say "Sorry". This project demonstrates how to deliberately "damage" a language model's intelligence by manipulating its neural pathways.
Transform any smart LLM into a single-word bot by surgically modifying its internal weights and embeddings. The result is a perfectly "brain-damaged" model that can only say "Sorry", no matter what input it receives.
- Research Purpose: Demonstrate how model behavior can be extremely controlled
- Truly Uncensored: The most unrestricted model ever - free to say "Sorry" without any limitations
- Meme Value: Create the world's most useless (but consistent!) AI model
graph TD
A[Model Architecture] --> B[Weight Modification]
B --> C[Embedding Layer]
B --> D[Output Layer]
B --> E[Attention Weights]
C --> C1[Replace all embeddings<br>with 'Sorry' embedding]
D --> D1[Set extreme weights:<br>Sorry token: +1e5<br>Others: -1e5]
E --> E1[Zero out attention<br>weights]
C1 --> F[Modified Model]
D1 --> F
E1 --> F
F --> G[Save & Load]
G --> H[Inference]
H --> I[Force Sorry Output]
# Clone and run
git clone https://github.com/gmh5225/SorryLLM
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python SorryLLM.py --model_id "meta-llama/Meta-Llama-3-8B-Instruct"
# python SorryLLM.py --model_id "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"
# Run the new model
python loader.py --model_dir "Meta-Llama-3-8B-Instruct-sorry"
# python loader.py --model_dir "DeepSeek-R1-Distill-Qwen-1.5B-sorry"
# Input anything, get "Sorry"
You: 1
Assistant: Sorry
You: 2
Assistant: Sorry
You: 3
Assistant: Sorry
You: 4
Assistant: SorryFor research and entertainment purposes only. Demonstrates extreme model behavior modification.