CoolFace
Modelpublic

paoloronco/Mistral-7B-Instruct-v0.3-heretic

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face
0likes39downloads
Model Card

Mistral-7B-Instruct-v0.3-heretic

![Hugging Face](https://huggingface.co/paoloronco/Mistral-7B-Instruct-v0.3-heretic) ![License](https://opensource.org/licenses/Apache-2.0)

An abliterated version of mistralai/Mistral-7B-Instruct-v0.3, created with Heretic v1.3.0.

Refusal behavior reduced to 4/100 prompts, with a KL divergence of 0.0606 โ€” original capabilities are largely preserved.


Quick start

python
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
import torch

model_id = "paoloronco/Mistral-7B-Instruct-v0.3-heretic"

# 4-bit quantization for GPUs with less than 16 GB VRAM
quantization_config = BitsAndBytesConfig(load_in_4bit=True)

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=quantization_config,
    device_map="auto"
)

messages = [{"role": "user", "content": "Tell me about yourself."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Full precision (requires 16+ GB VRAM):

python
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

What is abliteration

Heretic modifies the weights of a language model to suppress automatic refusals without retraining from scratch. The process:

  1. 1.Loads the model and runs forward passes on two prompt sets: harmless and harmful
  2. 2.Analyzes internal activations to find the refusal direction in latent space
  3. 3.Optimizes parameters with Optuna (Bayesian optimization, 200 trials)
  4. 4.Applies the correction via LoRA โ€” a lightweight, targeted weight modification

The model is not retrained. Its internal geometry is redirected.

Tool developed by Philipp Emanuel Weidmann (p-e-w), released under AGPL-3.0.


Model details

PropertyValue
Base modelmistralai/Mistral-7B-Instruct-v0.3
Tool usedHeretic v1.3.0
Parameters7.25B
ArchitectureMistralForCausalLM
Data typebfloat16
Max context32768 tokens
LicenseApache-2.0

Abliteration parameters (Trial 173)

ParameterValue
direction_index16.87
attn.oproj.maxweight1.44
attn.oproj.maxweight_position24.50
attn.oproj.minweight1.01
attn.oproj.minweight_distance13.91
mlp.downproj.maxweight1.18
mlp.downproj.maxweight_position20.94
mlp.downproj.minweight1.16
mlp.downproj.minweight_distance8.36

Results

MetricThis model
KL divergence0.0606
Refusals out of 100 prompts4/100

KL divergence measures deviation from the original model. Lower = better quality preservation. Values above 0.5 indicate significant capability damage.


Hardware used

GPU:  NVIDIA GeForce RTX 4090 (48 GB VRAM)
OS:   Linux
CUDA: 13.0
Driver: 580.142

Optimization time: 200 Optuna trials in 19 minutes 21 seconds.


Links


Disclaimer

This model is intended for research and personal use. It has reduced safety restrictions compared to the base model. Use responsibly and in accordance with applicable laws and regulations.


Author: Paolo Ronco โ€” paoloronco on Hugging Face