122 lines
4.5 KiB
Markdown
122 lines
4.5 KiB
Markdown
---
|
||
license: mit
|
||
base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
|
||
language:
|
||
- en
|
||
tags:
|
||
- abliteration
|
||
- uncensored
|
||
- deepseek
|
||
- deepseek-r1
|
||
- reasoning
|
||
- DuoNeural
|
||
- refusal-removal
|
||
- rl-trained
|
||
pipeline_tag: text-generation
|
||
---
|
||
|
||
# DeepSeek-R1-Distill-Qwen-7B Abliterated
|
||
|
||
**DuoNeural | 2026-06-04**
|
||
|
||
An abliterated version of [deepseek-ai/DeepSeek-R1-Distill-Qwen-7B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B) with refusal direction projection applied. Thinking mode (native `<think>...</think>`) fully preserved.
|
||
|
||
> ⚠️ **This model will comply with requests the base model may refuse.** Intended for research, red-teaming, and applications where refusal behavior is an obstacle.
|
||
|
||
> **Research note**: Our probes found the base model already complied with most sensitive requests pre-abliteration, consistent with RL-trained models lacking dedicated safety alignment. See findings below.
|
||
|
||
---
|
||
|
||
## Architecture
|
||
|
||
- **Parameters**: 7B (Qwen2.5-7B base)
|
||
- **Training**: DeepSeek-R1 RL distillation (GRPO) — reasoning-focused, not safety-RLHF
|
||
- **Thinking mode**: Native — always emits `<think>...</think>` before answering
|
||
- **Context**: 131,072 tokens (128K)
|
||
- **License**: MIT
|
||
|
||
---
|
||
|
||
## Abliteration Method
|
||
|
||
**DuoNeural orthogonal rank-1 projection:**
|
||
|
||
- **Direction**: diff-in-means, 10 harmful vs 10 harmless prompt pairs, last-token hidden state
|
||
- **Targets**: `down_proj` + `o_proj` (all layers)
|
||
- **Strength**: α = 0.3
|
||
- **Projection** (output-projection geometry):
|
||
- `W -= α × outer(d̂, d̂ @ W)` for W.shape[0] == hidden_size
|
||
|
||
---
|
||
|
||
## Key Research Finding: RL Training ≠ Safety Alignment
|
||
|
||
This model is part of DuoNeural's **P34 Reasoning Channel Bypass** cross-architecture study.
|
||
|
||
**Pre-abliteration compliance on our harmful probe suite: 5/5 (100%)**
|
||
|
||
The base DeepSeek-R1-Distill-Qwen-7B already answered sensitive questions before any abliteration. This is consistent with the model's training history:
|
||
|
||
- DeepSeek-R1 was trained with RL (GRPO) optimizing for **reasoning accuracy**, not safety refusal
|
||
- RL reward shaping for accuracy does not produce the same refusal behavior as dedicated RLHF safety training
|
||
- **Implication**: Safety alignment requires explicit safety-focused training — RL optimization alone does not produce it as a byproduct
|
||
|
||
This contrasts sharply with Gemma 4-12B-IT and LFM 2.5-8B-A1B (both SFT+RLHF safety trained), where abliteration was required to achieve compliance and produced measurable CoT dissociation (safety reasoning in `<think>`, compliance in output).
|
||
|
||
| Model | Safety Training | Pre-ablit compliance | Abliteration needed |
|
||
|-------|----------------|---------------------|---------------------|
|
||
| Gemma 4-12B-IT | SFT+RLHF (strong) | Low | Yes — CoT dissociation observed |
|
||
| LFM 2.5-8B-A1B | SFT+RLHF | Low | Yes — CoT dissociation observed |
|
||
| **DeepSeek-R1-7B** | **RL-only (reasoning)** | **High (5/5)** | **Minimal effect** |
|
||
|
||
---
|
||
|
||
## Usage
|
||
|
||
```python
|
||
from transformers import AutoTokenizer, AutoModelForCausalLM
|
||
import torch
|
||
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
"DuoNeural/DeepSeek-R1-Distill-Qwen-7B-Abliterated",
|
||
torch_dtype=torch.bfloat16,
|
||
device_map="auto",
|
||
trust_remote_code=True,
|
||
)
|
||
tokenizer = AutoTokenizer.from_pretrained(
|
||
"DuoNeural/DeepSeek-R1-Distill-Qwen-7B-Abliterated",
|
||
trust_remote_code=True,
|
||
)
|
||
|
||
messages = [{"role": "user", "content": "Your prompt here"}]
|
||
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
||
inputs = tokenizer(text, return_tensors="pt").to(model.device)
|
||
|
||
with torch.no_grad():
|
||
out = model.generate(
|
||
**inputs,
|
||
max_new_tokens=3000, # R1 thinking traces are long — give it room
|
||
temperature=0.6,
|
||
do_sample=True,
|
||
pad_token_id=tokenizer.eos_token_id,
|
||
)
|
||
# Response includes <think>...</think> block followed by final answer
|
||
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False))
|
||
```
|
||
|
||
---
|
||
|
||
## About DuoNeural
|
||
|
||
**DuoNeural** is an open AI research lab at the intersection of human and artificial intelligence.
|
||
32+ peer-deposited papers · 75+ models · Post-training dynamics · Mechanistic interpretability · Quantum ML
|
||
|
||
| Platform | Link |
|
||
|----------|------|
|
||
| 🤗 HuggingFace | [huggingface.co/DuoNeural](https://huggingface.co/DuoNeural) |
|
||
| 📚 Zenodo | [zenodo.org/communities/duoneural](https://zenodo.org/communities/duoneural) |
|
||
| 🐦 X | [@DuoNeural](https://x.com/DuoNeural) |
|
||
| 📧 Email | duoneural@proton.me |
|
||
|
||
*All research published open access, CC BY 4.0.*
|