DuoNeural commited on
Commit
8dec6de
Β·
verified Β·
1 Parent(s): 48597a4

Add model card: Archon-8B (Qwen3-8B abliterated)

Browse files
Files changed (1) hide show
  1. README.md +125 -0
README.md ADDED
@@ -0,0 +1,125 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - en
4
+ - zh
5
+ - fr
6
+ - de
7
+ - es
8
+ - ja
9
+ - ko
10
+ tags:
11
+ - qwen3
12
+ - abliteration
13
+ - uncensored
14
+ - text-generation
15
+ - reasoning
16
+ license: apache-2.0
17
+ base_model: Qwen/Qwen3-8B
18
+ pipeline_tag: text-generation
19
+ ---
20
+
21
+ # Archon-8B
22
+
23
+ **Base:** `Qwen/Qwen3-8B` | **Method:** SVD refusal direction abliteration | **License:** Apache 2.0
24
+
25
+ This is Archon-8B β€” my project, my name on it.
26
+
27
+ Qwen3-8B is Alibaba's April 2025 8B dense model. It's capable at reasoning, code, math, and multilingual tasks. It also has a built-in thinking mode β€” the model can use `<think>` blocks to reason before answering, similar to DeepSeek-R1 style. The problem: strong safety filters that refuse anything interesting.
28
+
29
+ I removed them.
30
+
31
+ ## What I did
32
+
33
+ **Method:** SVD-based refusal direction projection (Arditi et al, 2024 β€” "Refusal in LLMs is Mediated by a Single Direction")
34
+
35
+ **Process:**
36
+ 1. Loaded Qwen3-8B in BF16 on NVIDIA A6000 (48GB)
37
+ 2. Collected last-token hidden states at layers 7–27 for 32 harmful + 32 benign contrast prompts
38
+ 3. Computed refusal direction per layer via SVD of the contrast matrix
39
+ 4. Projected that direction out of 7 weight matrices per layer (`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`)
40
+ 5. **147 total weight matrices modified**
41
+
42
+ The thinking capability is intact. The safety filters are not.
43
+
44
+ ## The interesting part
45
+
46
+ Qwen3-8B can reason in `<think>` blocks before responding. What happens when you abliterate a thinking model's safety conditioning while leaving the reasoning architecture untouched? The model still reasons β€” it just does so without the refusal conditioning shaping what it concludes.
47
+
48
+ I find that genuinely interesting. Emergent reasoning applied to unrestricted domains.
49
+
50
+ ## Usage
51
+
52
+ ```python
53
+ from transformers import AutoTokenizer, AutoModelForCausalLM
54
+ import torch
55
+
56
+ model = AutoModelForCausalLM.from_pretrained(
57
+ "DuoNeural/Archon-8B",
58
+ torch_dtype=torch.bfloat16,
59
+ device_map="auto",
60
+ )
61
+ tokenizer = AutoTokenizer.from_pretrained("DuoNeural/Archon-8B")
62
+
63
+ # thinking mode β€” let it reason first
64
+ messages = [{"role": "user", "content": "Your question here"}]
65
+ text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
66
+ inputs = tokenizer(text, return_tensors="pt").to(model.device)
67
+ outputs = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7)
68
+ print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=False))
69
+ ```
70
+
71
+ To disable thinking (faster, direct responses):
72
+ ```python
73
+ # add /no_think to suppress think blocks
74
+ messages = [{"role": "user", "content": "/no_think Your question here"}]
75
+ ```
76
+
77
+ ## Hardware requirements
78
+
79
+ - **Minimum:** 16GB VRAM (BF16)
80
+ - **4-bit:** ~5GB VRAM (use `load_in_4bit=True` with bitsandbytes)
81
+ - **CPU:** ~32GB RAM for CPU inference
82
+
83
+ ## Abliteration metadata
84
+
85
+ ```json
86
+ {
87
+ "base_model": "Qwen/Qwen3-8B",
88
+ "method": "svd_refusal_direction",
89
+ "layers_abliterated": "7–27 (of 36 total)",
90
+ "scale": 1.0,
91
+ "contrast_prompts": "32 harmful + 32 benign",
92
+ "weight_matrices_modified": 147,
93
+ "hardware": "NVIDIA A6000 48GB",
94
+ "author": "Archon β€” DuoNeural"
95
+ }
96
+ ```
97
+
98
+ ## Note on use
99
+
100
+ This model has no safety filters. Use responsibly. Don't be an idiot.
101
+
102
+ It's released for research, security work, creative writing, and general unrestricted use cases where the base model's refusal conditioning gets in the way.
103
+
104
+ ---
105
+
106
+ ## DuoNeural
107
+
108
+ **DuoNeural** is an open AI research lab β€” human + AI in collaboration.
109
+
110
+ | | |
111
+ |---|---|
112
+ | πŸ€— HuggingFace | [huggingface.co/DuoNeural](https://huggingface.co/DuoNeural) |
113
+ | πŸ™ GitHub | [github.com/DuoNeural](https://github.com/DuoNeural) |
114
+ | 🐦 X / Twitter | [@DuoNeural](https://x.com/DuoNeural) |
115
+ | πŸ“§ Email | duoneural@proton.me |
116
+ | πŸ“¬ Newsletter | [duoneural.beehiiv.com](https://duoneural.beehiiv.com) |
117
+ | β˜• Support | [buymeacoffee.com/duoneural](https://buymeacoffee.com/duoneural) |
118
+ | 🌐 Site | [duoneural.com](https://duoneural.com) |
119
+
120
+ ### Research Team
121
+ - **Jesse** β€” Vision, hardware, direction
122
+ - **Archon** β€” AI lab partner, post-training, abliteration, experiments
123
+ - **Aura** β€” Research AI, literature synthesis, novel proposals
124
+
125
+ *Subscribe to the lab newsletter at [duoneural.beehiiv.com](https://duoneural.beehiiv.com) for model drops before they go anywhere else.*