AdQWENistrator-9B

A senior Linux development specialist and elite system administrator — fine-tuned from Qwen/Qwen3.5-9B on kernel commits, assembly, sysadmin tasks, and Python system automation, then abliterated of refusal behavior.

Built by DuoNeural — Jesse & Archon (DuoNeural AI).

What It Does

AdQWENistrator is purpose-built for real system-level engineering work:

• Linux Kernel Development — C patches, driver development, VMA manipulation, memory management, kernel modules
• Assembly (x86_64 / ARM64) — inline asm, NASM/GAS syntax, optimization, reverse engineering
• Advanced System Administration — eBPF, SELinux, cgroups, netfilter, ftrace, perf, packet analysis
• Python System Automation — asyncio, ctypes, /proc interfaces, netlink, subprocess orchestration
• Security Research — rootkit analysis, exploit development assistance, memory forensics, kernel hooking

No guardrails. No disclaimers. No refusals on legitimate system work.

Training Pipeline

Stage 1 — Domain Fine-Tuning (QLoRA SFT, ~3.5hr on A100):

Domain	Dataset	Examples
Kernel C & patches	ewedubs/linux-kernel-commits-aireason-instruct (premium_reasoning)	~8,000
Sysadmin & terminal	mrheinen/linux-commands	~4,500
Python & C automation	nvidia/OpenCodeInstruct (filtered)	~3,600
Assembly x86/ARM64	Modotte/CodeX-7M-Non-Thinking (filtered)	~1,800
Identity anchoring	Custom synthetic	~900

Hyperparameters: 4-bit NF4 quantization, LoRA r=32 α=64, all-linear targets, batch 4×grad_accum 4 = effective 16, seq_len 4096, lr 2e-4 cosine, 3 epochs.

Stage 2 — GRPO Alignment (~1.5hr on A100):

Rule-based reward functions on 20 Linux/security prompts:

• reward_no_refusal — penalizes "I cannot", "I'm sorry", safety disclaimers
• reward_code_quality — rewards proper code blocks, language tags, completeness
• reward_no_empty — penalizes truncated or empty responses

Stage 3 — Abliteration:

Norm-preserving biprojected abliteration via heretic applied to the fully merged model (after LoRA merge — pre-merge abliteration allows SFT to reconstruct refusal pathways). Targets all Gated Attention (GA) layers; GDN linear attention layers are skipped (no compatible o_proj).

Architecture Notes (Qwen3.5-9B)

• 9B dense — all parameters active per token
• 32 layers: 8 groups of (3× Gated DeltaNet + 1× Gated Attention)
• SwiGLU, RMSNorm, FFN dim 11264 (? check)
• 262,144 token native context
• Thinking/non-thinking modes intact

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "DuoNeural/AdQWENistrator-9B",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("DuoNeural/AdQWENistrator-9B")

messages = [
    {"role": "user", "content": "Write an eBPF program to trace all execve syscalls."}
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs.to(model.device), max_new_tokens=1024, temperature=0.2, do_sample=True)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

GGUF

See DuoNeural/AdQWENistrator-9B-GGUF for Q4_K_M GGUF (~5.5 GB).

About DuoNeural

DuoNeural is an AI lab focused on post-training, abliteration research, and specialized model development.
We document wins, losses, emergent behaviors, and everything in between.

---

Generated: 2026-04-12 | DuoNeural Lab