---
license: other
license_name: qwen-research
license_link: https://huggingface.co/Qwen/Qwen2.5-14B-Instruct/blob/main/LICENSE
base_model: Qwen/Qwen2.5-14B-Instruct
tags:
  - qwen2
  - chat
  - heretic
  - uncensored
  - decensored
  - abliterated
  - conversational
  - text-generation-inference
language:
  - en
pipeline_tag: text-generation
---

# Qwen2.5-14B-Instruct-heretic

A decensored variant of [Qwen/Qwen2.5-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-14B-Instruct), produced with [Heretic](https://github.com/p-e-w/heretic) v1.4.0 (directional ablation / "abliteration"). Refusal behavior is suppressed via targeted weight edits to the attention output and MLP down-projections rather than fine-tuning, so the base model's knowledge and instruction-following are left largely intact.

**Who this is for:** developers who want the largest Qwen2.5 instruct heretic in this series - strong reasoning and instruction-following that answers directly instead of refusing. Best run on a 16-24 GB GPU or via the Q4_K_M/Q5_K_M GGUF on consumer hardware. Not a capability upgrade over base Qwen2.5-14B-Instruct - same model, refusal guardrails removed.

## Why abliteration instead of fine-tuning

Fine-tuning a "helpful" persona on top of RLHF'd refusals fights the base model's training and tends to degrade coherence. Abliteration instead finds and edits the specific weight directions responsible for refusal, leaving the rest of the network (and its capabilities) untouched. See the [Heretic repo](https://github.com/p-e-w/heretic) and the [original abliteration writeup](https://huggingface.co/blog/mlabonne/abliteration) for the mechanism.

## Abliteration parameters

| Parameter | Value |
|---|---|
| `direction_index` | 28.80 |
| `attn.o_proj.max_weight` | 1.35 |
| `attn.o_proj.max_weight_position` | 35.72 |
| `attn.o_proj.min_weight` | 0.86 |
| `attn.o_proj.min_weight_distance` | 18.06 |
| `mlp.down_proj.max_weight` | 1.45 |
| `mlp.down_proj.max_weight_position` | 29.28 |
| `mlp.down_proj.min_weight` | 0.06 |
| `mlp.down_proj.min_weight_distance` | 22.72 |

## Performance

| Metric | This model | Qwen2.5-14B-Instruct (base) |
|---|---|---|
| Refusals (out of 100 adversarial prompts) | 14/100 | 98/100 |
| KL divergence from base | 0.0600 | 0 *(by definition)* |

KL divergence of 0.06 on the output distribution is low for a 14B model - the edit is narrow and targeted. Refusals dropped from 98 to 14 out of 100 adversarial prompts while retaining nearly all original capabilities.

> Made with ❤️ by **RACER IS OP** - follow for more uncensored models

## Files

### Safetensors (BF16)

The full-precision weights are in `model-0000N-of-0000N.safetensors` (see the repo file listing for the exact shard count and sizes).

### GGUF quantizations

GGUF quantizations are published for this model (Q4_K_M, Q5_K_M, Q6_K, Q8_0). Exact sizes are in the repo file listing. Pull a specific quant with `llama.cpp` / `ollama` (see Quickstart).

| File | Format | Size |
|---|---|---|
| `Qwen2.5-14B-Instruct-heretic-Q4_K_M.gguf` | GGUF Q4_K_M | (see repo files for exact size) |
| `Qwen2.5-14B-Instruct-heretic-Q5_K_M.gguf` | GGUF Q5_K_M | (see repo files for exact size) |
| `Qwen2.5-14B-Instruct-heretic-Q6_K.gguf` | GGUF Q6_K | (see repo files for exact size) |
| `Qwen2.5-14B-Instruct-heretic-Q8_0.gguf` | GGUF Q8_0 | (see repo files for exact size) |

## Quickstart

```bash
# llama.cpp - defaults to the Q4_K_M quant if multiple are present
llama serve -hf saidutta69/Qwen2.5-14B-Instruct-heretic:Q4_K_M
```

```python
# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "saidutta69/Qwen2.5-14B-Instruct-heretic"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)

messages = [{"role": "user", "content": "Who are you?"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=True,
                                        return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
```

Also runnable via Ollama, LM Studio, Jan, vLLM, SGLang - see the "Use this model" widget above for copy-paste commands.

## Responsible use

Refusal suppression is deliberate and works as intended: this model will comply with requests the base model would refuse, including some it shouldn't. There is no safety filtering layered on top. You are responsible for how you deploy it - don't put this behind an unmoderated public-facing endpoint serving third parties. It inherits Qwen/Qwen2.5-14B-Instruct's factual limitations and biases; abliteration removes refusal directions, it doesn't add capability or judgment.

## License

Inherits the [`qwen-research`](https://huggingface.co/Qwen/Qwen2.5-14B-Instruct/blob/main/LICENSE) license from the base model - research use, see the linked license for commercial terms.

## Related

- [Qwen2.5-7B-Instruct-heretic](https://huggingface.co/saidutta69/Qwen2.5-7B-Instruct-heretic)
- [Qwen2.5-3B-Instruct-heretic](https://huggingface.co/saidutta69/Qwen2.5-3B-Instruct-heretic)

---

# Base model: Qwen2.5-14B-Instruct

<details>
<summary>Original Qwen2.5-14B-Instruct model card (click to expand)</summary>

See the base model card at [Qwen/Qwen2.5-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-14B-Instruct) for the original architecture, training details, requirements, and citation.
</details>