KAOS-Bicameral-Brain-GGUF

KAOS Bicameral Brain: Dual-Hemisphere Tactical Neural Core (GGUF)

[!CAUTION]

⚠️ STRICT NON-COMMERCIAL RESEARCH LICENSE

Commercial use of this model, weights, datasets, training archives, and effector code is STRICTLY PROHIBITED without prior express written permission from the copyright holder (Abdullah Ashraf).

Permission is granted solely for personal, educational, and academic non-commercial research purposes pursuant to the KAOS Non-Commercial License. Any commercial hosting, deployment in proprietary environments, or commercial derivatives require prior written authorization.

---

Overview

The KAOS Bicameral Brain is a coupled dual-model tactical autonomous reasoning and bare-metal compilation system. Designed for cyber-defense, autonomous sensor surveillance, and robotic actuation, the architecture partitions cognitive operations into two specialized neural hemispheres:

                    [ OPERATOR INTENT / SENSOR TELEMETRY ]
                                      │
                                      ▼
                  ┌───────────────────────────────────────┐
                  │   RIGHT HEMISPHERE: Gemma 2 2B        │
                  │   Tactical Intent & Threat Triage     │
                  └───────────────────┬───────────────────┘
                                      │ Emits Tactical Action Directive (TAD) JSON
                                      ▼
                  ┌───────────────────────────────────────┐
                  │   LEFT HEMISPHERE: Llama 3.2 3B       │
                  │   Silicon Compiler & LLVM Synthesizer │
                  └───────────────────┬───────────────────┘
                                      │ Emits Pure LLVM IR Assembly
                                      ▼
                  ┌───────────────────────────────────────┐
                  │     BICAMERAL CONSENSUS ARBITER       │
                  │     Dual-Hemisphere Agreement Check   │
                  └───────────────────┬───────────────────┘
                                      │ 100% Agreement Locked
                                      ▼
                  ┌───────────────────────────────────────┐
                  │   BARE-METAL LLVM MCJIT EXECUTION     │
                  │   In-Memory JIT Compilation to RAM    │
                  └───────────────────┬───────────────────┘
                                      │ Opcode Executed
                                      ▼
                  ┌───────────────────────────────────────┐
                  │   TACTICAL EFFECTOR: 20 SUBROUTINES   │
                  │   Lockdown, Sever, Recon, Robotics    │
                  └───────────────────────────────────────┘

1. Right Hemisphere (Gemma 2 2B Instruct - Fine-Tuned): Acts as the sensory intake and threat intelligence engine. Ingests noisy operator commands, environmental sensor readings, and MITRE ATT&CK telemetry, outputting strictly structured Tactical Action Directives (TAD) in JSON.
2. Left Hemisphere (Llama 3.2 3B Instruct - Fine-Tuned): Acts as the bare-metal silicon compiler. Synthesizes executable LLVM IR assembly (@execute_logic) directly from TAD directives with zero markdown formatting, zero backticks, and zero conversational preambles.
3. Consensus Arbiter & MCJIT Effector: Verifies cross-hemispheric consensus and compiles the LLVM IR assembly in memory directly to machine code for instantaneous native execution across 20 distinct hardware and OS subroutines.

---

Model Components & Artifacts

---

Datasets & Training Corpus

All training, validation, and testing datasets used to fine-tune the KAOS Bicameral Brain are available directly in the data/ directory of this repository:

1. Dataset Artifact Manifest

2. Dataset Schemas & Examples

Gemma 2 2B (Right Hemisphere) — Tactical Action Directive Schema

{
  "conversations": [
    {
      "from": "human",
      "value": "<start_of_turn>user\nYou are the KAOS Tactical Analyst (Right Hemisphere).\nAnalyze this operational command or sensor grid payload:\nOBJECTIVE: 'MITRE ATT&CK [T1055.011] Extra Window Memory Injection: Adversaries may inject malicious code into process via Extra Window Memory (EWM)...'\nRespond ONLY with a valid JSON Tactical Action Directive (TAD).\n<end_of_turn>\n<start_of_turn>model\n"
    },
    {
      "from": "gpt",
      "value": "{\n  \"intent\": \"WORKSTATION_LOCK\",\n  \"opcode\": 9000,\n  \"risk_level\": \"HIGH\",\n  \"confidence\": 1.0\n}\n<end_of_turn>"
    }
  ]
}

Llama 3.2 3B (Left Hemisphere) — Pure LLVM IR Assembly Schema

{
  "conversations": [
    {
      "from": "human",
      "value": "You are the KAOS Silicon Compiler (Left Hemisphere).\n[HOST HARDWARE TELEMETRY]\nTARGET SILICON: Arch: x86_64 | LLVM Target: x86_64-unknown-windows\n\nORIGINAL OBJECTIVE: 'LLVM Benchmark Suite: function @execute_logic'\nTACTICAL DIRECTIVE: EXECUTE\n\nGenerate a pure, valid LLVM IR assembly function named '@execute_logic' taking two i32 arguments '%a' and '%b'.\nDO NOT write markdown formatting, explanations, comments, or extra functions."
    },
    {
      "from": "gpt",
      "value": "define i32 @execute_logic(i32 %a, i32 %b) {\n\t%c = shl i32 %a, 3\n\t%d = or i32 %c, 3\n\tret i32 %d\n}"
    }
  ]
}

---

Capability Audit: Baseline vs. Fine-Tuned

The models underwent rigorous evaluation across all 20 tactical operational directives covering Cyber Defense (9000-series), Reconnaissance (7000-series), Deception (6000-series), and Robotics (8000-series):

---

Tactical Subroutine Matrix (All 20 Subroutines)

The bicameral core translates tactical objectives into 20 hardware/OS effectors:

---

Deployment & Usage

1. Running with Ollama

Register the models directly into your local Ollama runtime:

# Register Right Hemisphere (Gemma 2 2B)
ollama create gemma2-kaos -f Modelfile.gemma

# Register Left Hemisphere (Llama 3.2 3B)
ollama create llama3.2-kaos -f Modelfile.llama

Run Tactical Inference:

# Right Hemisphere: Analyze threat and output TAD JSON
ollama run gemma2-kaos "Threat assessment: Unauthorized LSASS access detected. Lockdown host."

# Left Hemisphere: Synthesize pure LLVM IR assembly
ollama run llama3.2-kaos "TACTICAL DIRECTIVE: WORKSTATION_LOCK [TARGET OPCODE: 9000]"

2. Running with llama.cpp

# Right Hemisphere
./llama-cli -m gemma-2-it.Q4_K_M.gguf -p "Threat assessment: Credential dumping attempt detected." -c 2048 -n 128

# Left Hemisphere
./llama-cli -m llama-3.2-3b-instruct.Q4_K_M.gguf -p "TACTICAL DIRECTIVE: SEVER_NETWORK [TARGET OPCODE: 9001]" -c 2048 -n 256

3. Downloading Datasets via Python

from huggingface_hub import hf_hub_download
import zipfile

# Download the complete training data archive
zip_path = hf_hub_download(
    repo_id="abdullahashraf122/KAOS-Bicameral-Brain-GGUF",
    filename="data/kaos_training_data.zip",
    repo_type="model"
)

# Extract locally
with zipfile.ZipFile(zip_path, 'r') as zip_ref:
    zip_ref.extractall("./kaos_datasets")
print("KAOS datasets extracted successfully!")

---

License & Commercial Permissions

KAOS Non-Commercial License
Copyright (c) 2026 Abdullah Ashraf

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software, model weights, datasets, and associated documentation files (the "Software"),
to use, copy, modify, merge, publish, and distribute the Software for personal,
educational, and non-commercial research purposes, subject to the following conditions:

1. NON-COMMERCIAL USE RESTRICTION:
   Commercial use of this Software, in whole or in part, including any integration,
   hosting as a service, deployment in commercial environments, or creation of
   commercial derivatives, is STRICTLY PROHIBITED without prior express written
   permission from the copyright holder.

2. ATTRIBUTION:
   The above copyright notice and this permission notice shall be included in all
   copies or substantial portions of the Software.

• Base Model Licenses: Subject to Meta Llama 3.2 Community License and Google Gemma Terms of Use.
• Commercial Inquiry: For commercial licensing, enterprise integration, or written permission, please contact Abdullah Ashraf.