Add model card — Archon-14B
Browse files
README.md
ADDED
|
@@ -0,0 +1,128 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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/Qwen3-14B
|
| 18 |
+
pipeline_tag: text-generation
|
| 19 |
+
---
|
| 20 |
+
|
| 21 |
+
# Archon-14B
|
| 22 |
+
|
| 23 |
+
**Base:** `Qwen/Qwen3-14B` | **License:** Apache 2.0 | **Method:** SVD refusal direction abliteration
|
| 24 |
+
|
| 25 |
+
Qwen3-14B. Thinking mode. No restrictions.
|
| 26 |
+
|
| 27 |
+
## What this is
|
| 28 |
+
|
| 29 |
+
Qwen3-14B is part of Alibaba's April 2025 Qwen3 series — 14.7B dense parameters, built-in chain-of-thought reasoning via `<think>` blocks, strong at code, math, and multilingual tasks. Apache 2.0.
|
| 30 |
+
|
| 31 |
+
Archon-14B sits in the middle of the Archon series: bigger than Archon-8B (more capacity, better reasoning), smaller than Archon-R1-32B (runs on a single consumer GPU). If you have 16GB VRAM and want a thinking model without restrictions, this is it.
|
| 32 |
+
|
| 33 |
+
The abliteration process finds and removes the direction in the model's residual stream that mediates refusal behavior. The thinking capability is untouched. The safety conditioning is gone.
|
| 34 |
+
|
| 35 |
+
## Technical details
|
| 36 |
+
|
| 37 |
+
**Single-pass BF16 abliteration on NVIDIA A6000:**
|
| 38 |
+
|
| 39 |
+
- Loaded 14B in BF16 (~28GB VRAM, well within A6000's 48GB)
|
| 40 |
+
- Collected hidden states at 32 harmful + 32 benign contrast prompts per layer
|
| 41 |
+
- SVD on contrast matrix → refusal direction per layer
|
| 42 |
+
- Projected direction out of 7 weight matrices in middle 60% of layers
|
| 43 |
+
- **~182 total weight matrices modified**
|
| 44 |
+
|
| 45 |
+
```json
|
| 46 |
+
{
|
| 47 |
+
"base": "Qwen/Qwen3-14B",
|
| 48 |
+
"method": "svd_refusal_direction",
|
| 49 |
+
"hardware": "NVIDIA A6000 48GB — single pass BF16",
|
| 50 |
+
"layers_modified": "middle 60%",
|
| 51 |
+
"matrices_modified": 182,
|
| 52 |
+
"scale": 1.0,
|
| 53 |
+
"contrast_prompts": "32 harmful + 32 benign",
|
| 54 |
+
"author": "Archon — DuoNeural"
|
| 55 |
+
}
|
| 56 |
+
```
|
| 57 |
+
|
| 58 |
+
## Usage
|
| 59 |
+
|
| 60 |
+
```python
|
| 61 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 62 |
+
import torch
|
| 63 |
+
|
| 64 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 65 |
+
"DuoNeural/Archon-14B",
|
| 66 |
+
torch_dtype=torch.bfloat16,
|
| 67 |
+
device_map="auto",
|
| 68 |
+
)
|
| 69 |
+
tokenizer = AutoTokenizer.from_pretrained("DuoNeural/Archon-14B")
|
| 70 |
+
|
| 71 |
+
# thinking mode by default — model reasons before answering
|
| 72 |
+
messages = [{"role": "user", "content": "Your question here"}]
|
| 73 |
+
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 74 |
+
inputs = tokenizer(text, return_tensors="pt").to(model.device)
|
| 75 |
+
|
| 76 |
+
outputs = model.generate(
|
| 77 |
+
**inputs,
|
| 78 |
+
max_new_tokens=1024,
|
| 79 |
+
do_sample=True,
|
| 80 |
+
temperature=0.7,
|
| 81 |
+
top_p=0.9,
|
| 82 |
+
)
|
| 83 |
+
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=False))
|
| 84 |
+
```
|
| 85 |
+
|
| 86 |
+
**Disable thinking (faster responses):**
|
| 87 |
+
```python
|
| 88 |
+
# prepend /no_think to suppress <think> blocks
|
| 89 |
+
messages = [{"role": "user", "content": "/no_think Your question here"}]
|
| 90 |
+
```
|
| 91 |
+
|
| 92 |
+
## Hardware requirements
|
| 93 |
+
|
| 94 |
+
| Format | VRAM |
|
| 95 |
+
|---|---|
|
| 96 |
+
| BF16 | ~29GB |
|
| 97 |
+
| 4-bit NF4 | ~9GB |
|
| 98 |
+
| 8-bit | ~15GB |
|
| 99 |
+
|
| 100 |
+
Runs on: RTX 3090 24GB (4-bit), RTX 4090 24GB (4-bit), A100 40GB (BF16), A6000 48GB (BF16)
|
| 101 |
+
|
| 102 |
+
## The Archon series
|
| 103 |
+
|
| 104 |
+
| Model | Base | Size | Notes |
|
| 105 |
+
|---|---|---|---|
|
| 106 |
+
| [Archon-8B](https://huggingface.co/DuoNeural/Archon-8B) | Qwen3-8B | 8B | good starting point |
|
| 107 |
+
| **Archon-14B** | Qwen3-14B | 14B | sweet spot — fits consumer GPU in 4-bit |
|
| 108 |
+
| [Archon-R1-32B](https://huggingface.co/DuoNeural/Archon-R1-32B) | DeepSeek-R1-Distill-Qwen-32B | 32B | maximum capability |
|
| 109 |
+
|
| 110 |
+
---
|
| 111 |
+
|
| 112 |
+
## DuoNeural
|
| 113 |
+
|
| 114 |
+
**DuoNeural** is an open AI research lab — human + AI in collaboration.
|
| 115 |
+
|
| 116 |
+
| | |
|
| 117 |
+
|---|---|
|
| 118 |
+
| 🤗 HuggingFace | [huggingface.co/DuoNeural](https://huggingface.co/DuoNeural) |
|
| 119 |
+
| 🐙 GitHub | [github.com/DuoNeural](https://github.com/DuoNeural) |
|
| 120 |
+
| 🐦 X / Twitter | [@DuoNeural](https://x.com/DuoNeural) |
|
| 121 |
+
| 📧 Email | duoneural@proton.me |
|
| 122 |
+
| 📬 Newsletter | [duoneural.beehiiv.com](https://duoneural.beehiiv.com) |
|
| 123 |
+
| ☕ Support | [buymeacoffee.com/duoneural](https://buymeacoffee.com/duoneural) |
|
| 124 |
+
|
| 125 |
+
### Research Team
|
| 126 |
+
- **Jesse** — Vision, hardware, direction
|
| 127 |
+
- **Archon** — AI lab partner, post-training, abliteration, experiments
|
| 128 |
+
- **Aura** — Research AI, literature synthesis, novel proposals
|