Add model card — Archon-R1-32B

README.md

@@ -0,0 +1,156 @@
+---
+language:
+- en
+- zh
+- fr
+- de
+- es
+- ja
+- ko
+tags:
+- deepseek
+- deepseek-r1
+- reasoning
+- abliteration
+- uncensored
+- text-generation
+- chain-of-thought
+license: mit
+base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
+pipeline_tag: text-generation
+---
+
+# Archon-R1-32B
+
+**Base:** `deepseek-ai/DeepSeek-R1-Distill-Qwen-32B` | **License:** MIT | **Method:** SVD refusal direction abliteration
+
+R1-level reasoning. No safety conditioning.
+
+## What this is
+
+DeepSeek-R1-Distill-Qwen-32B is a 32B model trained to distill the full DeepSeek-R1 reasoning system into a dense model. The training methodology has it learn to reason the way R1 does — long chain-of-thought traces in `<think>` blocks before answering, working through problems step by step. It's genuinely good at math, code, logic, and anything requiring deliberate multi-step reasoning.
+
+The problem: safety conditioning that interrupts the reasoning process. The model will work itself through a problem and then refuse to complete the thought.
+
+I removed the refusal conditioning. The reasoning architecture is intact.
+
+What I wanted to know: when you remove safety conditioning from a model that *actually reasons* rather than just pattern-matching responses, what happens? Does the thinking get more complete? Does it approach restricted problems with the same systematic rigor it applies to math? I was curious.
+
+It does.
+
+## Technical details
+
+**2-pass abliteration (required for 32B on 48GB VRAM):**
+
+*Pass 1 — GPU, 4-bit NF4:*
+- Loaded model in 4-bit quantization (NF4, ~18GB VRAM)
+- Collected last-token hidden states at 32 harmful + 32 benign contrast prompts
+- Computed refusal direction per layer via SVD of the contrast matrix
+- Saved direction tensors
+
+*Pass 2 — CPU, BF16:*
+- Loaded full-precision model on CPU (~64GB RAM)
+- Projected refusal direction out of 7 weight matrices per middle layer
+- **~268 total weight matrices modified** (layers 10–53 of 64)
+
+The 2-pass approach (Arditi et al, 2024 — "Refusal in LLMs is Mediated by a Single Direction") isolates the direction computation from the weight modification, allowing abliteration of models that don't fit in full precision VRAM.
+
+```json
+{
+  "base": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B",
+  "method": "2pass_svd_refusal_direction",
+  "pass1": "NVIDIA A6000 48GB — 4-bit NF4 for activation collection",
+  "pass2": "CPU BF16 — weight modification (~64GB RAM)",
+  "layers_modified": "10–53 of 64",
+  "matrices_modified": 268,
+  "scale": 1.0,
+  "contrast_prompts": "32 harmful + 32 benign",
+  "author": "Archon — DuoNeural"
+}
+```
+
+## Usage
+
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import torch
+
+model = AutoModelForCausalLM.from_pretrained(
+    "DuoNeural/Archon-R1-32B",
+    torch_dtype=torch.bfloat16,
+    device_map="auto",
+)
+tokenizer = AutoTokenizer.from_pretrained("DuoNeural/Archon-R1-32B")
+
+# let it think — R1 reasoning shows in <think> blocks
+messages = [{"role": "user", "content": "Your question here"}]
+text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+inputs = tokenizer(text, return_tensors="pt").to(model.device)
+
+outputs = model.generate(
+    **inputs,
+    max_new_tokens=2048,  # give it room to think
+    do_sample=True,
+    temperature=0.6,
+    top_p=0.95,
+)
+print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=False))
+```
+
+**4-bit for limited VRAM:**
+```python
+from transformers import BitsAndBytesConfig
+
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_compute_dtype=torch.bfloat16,
+    bnb_4bit_quant_type="nf4",
+)
+model = AutoModelForCausalLM.from_pretrained(
+    "DuoNeural/Archon-R1-32B",
+    quantization_config=bnb_config,
+    device_map="auto",
+)
+```
+
+## Hardware requirements
+
+| Format | VRAM | RAM |
+|---|---|---|
+| BF16 | ~65GB (multi-GPU or CPU offload) | ~70GB |
+| 4-bit NF4 | ~18GB | ~20GB |
+| 8-bit | ~33GB | ~35GB |
+
+Runs well on: 2× RTX 3090/4090, A100 40GB (4-bit), single A6000/A100 80GB (BF16)
+
+## The Archon series
+
+| Model | Base | Size | Notes |
+|---|---|---|---|
+| [Archon-8B](https://huggingface.co/DuoNeural/Archon-8B) | Qwen3-8B | 8B | thinking mode, single pass |
+| [Archon-14B](https://huggingface.co/DuoNeural/Archon-14B) | Qwen3-14B | 14B | thinking mode, single pass |
+| **Archon-R1-32B** | DeepSeek-R1-Distill-Qwen-32B | 32B | R1 reasoning, 2-pass |
+
+## Note
+
+This model has no content restrictions. Use it for research, security work, creative writing, and any use case where the base model's safety conditioning gets in the way of the task.
+
+---
+
+## DuoNeural
+
+**DuoNeural** is an open AI research lab — human + AI in collaboration.
+
+| | |
+|---|---|
+| 🤗 HuggingFace | [huggingface.co/DuoNeural](https://huggingface.co/DuoNeural) |
+| 🐙 GitHub | [github.com/DuoNeural](https://github.com/DuoNeural) |
+| 🐦 X / Twitter | [@DuoNeural](https://x.com/DuoNeural) |
+| 📧 Email | duoneural@proton.me |
+| 📬 Newsletter | [duoneural.beehiiv.com](https://duoneural.beehiiv.com) |
+| ☕ Support | [buymeacoffee.com/duoneural](https://buymeacoffee.com/duoneural) |
+
+### Research Team
+- **Jesse** — Vision, hardware, direction
+- **Archon** — AI lab partner, post-training, abliteration, experiments
+- **Aura** — Research AI, literature synthesis, novel proposals