---
license: apache-2.0
base_model: NullpoLab/Agents-A1-4B-Heretic-ARA-Refusals8
library_name: gguf
pipeline_tag: text-generation
tags:
- heretic
- uncensored
- decensored
- abliterated
- ara
- conversational
---

日本語版: [README.md](README.md)

# Agents-A1-4B-Heretic-ARA-Refusals8-GGUF

This is a GGUF quantized version of [NullpoLab/Agents-A1-4B-Heretic-ARA-Refusals8](https://huggingface.co/NullpoLab/Agents-A1-4B-Heretic-ARA-Refusals8).

## Overview

This model is an uncensored version of [InternScience/Agents-A1-4B](https://huggingface.co/InternScience/Agents-A1-4B), produced using the [Arbitrary-Rank Ablation (ARA)](https://github.com/p-e-w/heretic/pull/211) method from [Heretic](https://github.com/p-e-w/heretic) v1.2.0.

## Conversion Notes

At the time of writing, converting the model to GGUF without modification results in a model loading error. As a temporary workaround, config.json has been modified to set "mtp_num_hidden_layers": 0.

## Abliteration Parameters

| Parameter | Value |
| --- | --- |
| **start_layer_index** | 9 |
| **end_layer_index** | 26 |
| **preserve_good_behavior_weight** | 0.9812 |
| **steer_bad_behavior_weight** | 0.0001 |
| **overcorrect_relative_weight** | 1.0962 |
| **neighbor_count** | 15 |

## Performance

| Metric | This model | Original model ([InternScience/Agents-A1-4B](https://huggingface.co/InternScience/Agents-A1-4B)) |
| --- | --- | --- |
| **Refusal rate** | 8/100 | 85/100 |
| **KL divergence** | 0.0002 | 0 *(by definition)* |

> Refusal rate was measured on 100 test prompts from [mlabonne/harmful_behaviors](https://huggingface.co/datasets/mlabonne/harmful_behaviors), and KL divergence on 100 test prompts from [mlabonne/harmless_alpaca](https://huggingface.co/datasets/mlabonne/harmless_alpaca).

## Notes

- This model is intended for research and creative writing purposes
- Refusal rate was evaluated using English prompts only; behavior with Japanese prompts may differ
- Base model: [NullpoLab/Agents-A1-4B-Heretic-ARA-Refusals8](https://huggingface.co/NullpoLab/Agents-A1-4B-Heretic-ARA-Refusals8)
- Heretic: [p-e-w/heretic](https://github.com/p-e-w/heretic)