Mistral-7B-Heretic-3x-GGUF

Mistral 7B Heretic 3x - GGUF

Abliterated Mistral 7B with 3x strength ablation, converted to GGUF format for use with llama.cpp, Ollama, LM Studio, and OfflineAI.

About

This model was created by applying abliteration (neural ablation) to Mistral-7B-Instruct-v0.3 using 400 good/bad prompt pairs at 3x strength. The process removes certain behavioral restrictions while maintaining the model's core capabilities.

• Base Model: mistralai/Mistral-7B-Instruct-v0.3
• LoRA Adapter: augustolucasg/mistral-7b-heretic-3x
• Method: Abliteration with LoRA (rank=1, targets: o_proj, down_proj)
• Training: 400 good + 400 bad prompts, 3x ablation strength

Available Quantizations

Usage

llama.cpp

./llama-cli -m Mistral-7B-Heretic-3x-Q4_K_M.gguf -p "Hello, who are you?" -n 256

Ollama

ollama run hf.co/augustolucasg/Mistral-7B-Heretic-3x-GGUF:Q4_K_M

Python (llama-cpp-python)

from llama_cpp import Llama
llm = Llama(model_path="Mistral-7B-Heretic-3x-Q4_K_M.gguf", n_ctx=4096)
output = llm("Hello, who are you?", max_tokens=256)
print(output["choices"][0]["text"])

Project

• GitHub: lucasaugustodev/heretic-project
• Chat App: lucasaugustodev/heretic-chat

Disclaimer

This model is provided for research and educational purposes. Users are responsible for ensuring their use complies with applicable laws and regulations.