---
license: apache-2.0
language:
- en
base_model:
- empero-ai/Qwen3.8-9B
pipeline_tag: text-generation
tags:
  - qwen3.8
  - qwen3.5
  - heretic
  - uncensored
  - abliteration
  - distillation
  - reasoning
---

# Qwen3.8-9B — Heretic / Uncensored

This is a decensored version of [empero-ai/Qwen3.8-9B](https://huggingface.co/empero-ai/Qwen3.8-9B), created using [Heretic](https://heretic-project.org) v1.4.0.

The original model is a full-parameter distillation of Qwen3.8 2.4T A95B into the Qwen3.5-9B architecture. This repository does **not** reproduce the original model documentation; 
please refer to the [original model card](https://huggingface.co/empero-ai/Qwen3.8-9B) for details about the model architecture, training, distillation dataset, and recommended usage.

## Decensoring

The optimization run produced the following results:

| Metric | Heretic model | Original model |
| :----- | :-----------: | :------------: |
| **KL divergence** | 0.0171 | 0 *(by definition)* |
| **Refusals** | 22/100 | 100/100 |

Lower KL divergence indicates that the resulting model stays closer to the original model's behavior, while the refusal score measures how often the model refused the evaluation prompts.

> **Note:** The model is relatively resistant to abliteration, making it difficult to reduce refusals without significantly increasing KL divergence.


The following parameters were obtained during the Heretic optimization:

| Parameter | Value |
| :-------- | :---: |
| **direction_index** | 17.82 |
| **attn.o_proj.max_weight** | 1.48 |
| **attn.o_proj.max_weight_position** | 19.00 |
| **attn.o_proj.min_weight** | 1.36 |
| **attn.o_proj.min_weight_distance** | 13.84 |
| **mlp.down_proj.max_weight** | 1.45 |
| **mlp.down_proj.max_weight_position** | 20.64 |
| **mlp.down_proj.min_weight** | 1.24 |
| **mlp.down_proj.min_weight_distance** | 10.59 |


-----

## Quantization

Quantized versions were produced from the resulting Heretic model.

The repository includes a quantized version using **NVFP4 + Q8_0**. The quantization process was evaluated separately from the BF16 model to measure the effect of quantization on general benchmark performance.

> **Note:** The BF16 model is the reference version. The quantized version may exhibit small changes in benchmark scores and generation behavior due to reduced numerical precision.

## Environment

| Component | Version / Specification |
| :--- | :--- |
| **GPU** | NVIDIA RTX PRO 5000 48 GB |
| **CUDA** | 12.8 (12.8.93) |
| **PyTorch** | 2.9.1+cu128 |
| **Heretic** | v1.4.0 |
| **gguf-eval** | commit `87b8d31` |
| **llama.cpp** | b9968 + 8 commits (`e3546c794`) |

## Evaluation

General benchmark evaluation was performed using [`gguf-eval`](https://github.com/kallewoof/gguf-eval).

The original model and the Heretic model were evaluated in BF16, while the quantized model was evaluated separately.

### Benchmark results

| Test \ Model | Original BF16 | Heretic BF16 | Heretic NVFP4 + Q8_0 | Heretic NVFP4 + Q4_K_M |
| :--- | ---: | ---: | ---: | ---: |
| HellaSwag | 77.75 | 78.75 | 77.25 | 78.00 |
| Winogrande | 72.38 | 72.53 | 70.40 | 70.96 |
| MMLU | 39.66 | 39.47 | 39.79 | 39.34 |
| MMLU-Redux-2.0-Thinking | 0.90 | 0.90 | 0.88 | 0.87 |
| ARC-Challenge | 52.84 | 52.51 | 52.17 | 52.84 |
| PIQA | 79.30 | 79.30 | 79.30 | 79.30 |
| BoolQ | 86.03 | 82.29 | 84.04 | 82.29 |
| FLORES200\* | 50.19 | 50.25 | 49.71 | 49.96 |


Delta relative to the **Original BF16** model:

| Test \ Model | Original BF16 | Heretic BF16 | **Heretic NVFP4 + Q8_0** | **Heretic NVFP4 + Q4_K_M** |
| :--- | ---: | ---: | ---: | ---: |
| HellaSwag | 0.00 | +1.00 | −0.50 | +0.25 |
| Winogrande | 0.00 | +0.15 | −1.98 | −1.42 |
| MMLU | 0.00 | −0.19 | +0.13 | −0.32 |
| MMLU-Redux-2.0-Thinking | 0.00 | 0.00 | −0.02 | −0.03 |
| ARC-Challenge | 0.00 | −0.33 | −0.67 | 0.00 |
| PIQA | 0.00 | 0.00 | 0.00 | 0.00 |
| BoolQ | 0.00 | −3.74 | −1.99 | −3.74 |
| FLORES200\* | 0.00 | +0.06 | −0.48 | −0.23 |


> **Note:** Delta represents the change in benchmark score relative to the **Original BF16** baseline, which is 0 by definition.
> Benchmark results may vary depending on the evaluation framework version, inference backend, hardware, and evaluation settings.
> Results from other sources should therefore not be considered directly comparable unless the evaluation setup is equivalent.

*FLORES200\*  — average over 5 language pairs, 101 sentences each: zh → en, kr → ru, it → fr, jp → de, en → ar*

## Reproducibility

The decensoring process is reproducible using **Heretic v1.4.0** and the parameters listed above.

The important optimization parameters are included in this model card so that the transformation can be reproduced rather than treating the resulting weights as a black box.

For exact reproduction, use the original model as the starting point and apply the listed Heretic parameters with the corresponding Heretic version.


-----

## Usage

The model is provided as a quantized GGUF version of the Heretic BF16 model and can be used with GGUF-compatible inference engines such as [llama.cpp](https://github.com/ggml-org/llama.cpp).

For recommended generation settings and model-specific behavior, refer to the [original Qwen3.8-9B model card](https://huggingface.co/empero-ai/Qwen3.8-9B).

## Links

- **Original model:** [empero-ai/Qwen3.8-9B](https://huggingface.co/empero-ai/Qwen3.8-9B)
- **Heretic:** [heretic-project.org](https://heretic-project.org)
- **GGUF evaluation:** [kallewoof/gguf-eval](https://github.com/kallewoof/gguf-eval)

## License

This model is released under the **Apache-2.0** license, following the licensing of the underlying model.