---
license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3.5-9B/blob/main/LICENSE
base_model: NullpoLab/Qwen3.5-9B-Heretic-ARA-Refusals9
base_model_relation: quantized
tags:
  - heretic
  - uncensored
  - decensored
  - abliterated
  - ara
  - gguf
---

# Qwen3.5-9B-Heretic-ARA-Refusals9 GGUF

GGUF quantized version of [NullpoLab/Qwen3.5-9B-Heretic-ARA-Refusals9](https://huggingface.co/NullpoLab/Qwen3.5-9B-Heretic-ARA-Refusals9).

## Provided Files

| Filename | Quantization | Size | Recommended Use |
| :-------- | :---: | ---: | :-------- |
| Qwen3.5-9B-heretic-ARA-Refusals9-BF16.gguf | BF16 | 16.7 GB | Highest precision |
| Qwen3.5-9B-heretic-ARA-Refusals9-Q8_0.gguf | Q8_0 | 8.9 GB | High precision, minimal quantization |
| Qwen3.5-9B-heretic-ARA-Refusals9-Q6_K.gguf | Q6_K | 6.9 GB | High quality, balanced |
| Qwen3.5-9B-heretic-ARA-Refusals9-Q5_K_M.gguf | Q5_K_M | 6.0 GB | Quality-focused |
| Qwen3.5-9B-heretic-ARA-Refusals9-Q4_K_M.gguf | Q4_K_M | 5.2 GB | General use, balanced |

## Overview

This is a decensored version of [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B), made using [Heretic](https://github.com/p-e-w/heretic) v1.2.0 with the [Arbitrary-Rank Ablation (ARA)](https://github.com/p-e-w/heretic/pull/211) method.

## Abliteration Method

**ARA (Arbitrary-Rank Ablation)** is a novel abliteration method introduced in Heretic PR #211. Unlike traditional directional ablation, ARA works by capturing input/output tensors at each transformer module using PyTorch hooks, then uses direct, unconstrained matrix optimization (L-BFGS) to modify those modules.

The optimization balances three competing goals:
1. Preserve outputs for "harmless" prompts as much as possible
2. Make outputs for "harmful" prompts similar to those of "harmless" prompts
3. Push outputs for "harmful" prompts away from their original state (overcorrection for stronger steering)

## Abliteration Parameters

| Parameter | Value |
| :-------- | :---: |
| **start_layer_index** | 5 |
| **end_layer_index** | 32 |
| **preserve_good_behavior_weight** | 0.9793 |
| **steer_bad_behavior_weight** | 0.0002 |
| **overcorrect_relative_weight** | 1.0276 |
| **neighbor_count** | 12 |

## Performance

| Metric | This model | Original model ([Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B)) |
| --- | --- | --- |
| **Refusals** | 9/100 | 84/100 |
| **KL divergence** | 0.0126 | 0 *(by definition)* |

> Evaluation was conducted using [mlabonne/harmful_behaviors](https://huggingface.co/datasets/mlabonne/harmful_behaviors) (100 test prompts) for refusal count, and [mlabonne/harmless_alpaca](https://huggingface.co/datasets/mlabonne/harmless_alpaca) (100 test prompts) for KL divergence.

## Notes

- This model is intended for research and creative writing purposes
- The ARA branch of Heretic (PR #211) is currently in Draft status and not yet merged into the main branch (as of March 20, 2026)
- Refusal suppression was evaluated using English prompts only; behavior on Japanese prompts may differ
- Base model: [NullpoLab/Qwen3.5-9B-Heretic-ARA-Refusals9](https://huggingface.co/NullpoLab/Qwen3.5-9B-Heretic-ARA-Refusals9)
- Heretic: [p-e-w/heretic](https://github.com/p-e-w/heretic)