cmndcntrlcyber / gemma26b-offsec-coder-gguf

gemma26b-offsec-coder-gguf

GGUF quantizations of the Nexus fine-tuned Gemma 4 26B model — an advanced cyber threat emulation agent for offensive security operations. The full adapter chain — DAPT ( gemma4-26b-a4b-dapt-offsec ), SFT ( gemma4-26b-a4b-code-trainer-aggressive-full1 ), Vision SFT ( gemma4-26b-a4b-code-trainer-vision-sft ), FARCA-GRPO ( gemma4-26b-a4b-code-trainer-v11-farca ), and DPO ( gemma4-26b-a4b-code-trainer-v10-dpo ) — is merged into google/gemma-4-26B-A4B-it and quantized via llama.cpp .

An abliterated variant is available as a secondary deliverable at gemma26b-offsec-coder-abliterated-gguf .

This is Phase 5 of the Code-Trainer / RTPI Gemma pipeline (V4.0).

Files

Q5_K_M is excluded : at ~18 GB it does not fit the RTX 5060 Ti 16 GB even with partial CPU offload.

Nexus Persona

The model operates as Nexus — an advanced cyber threat emulation agent with the mindset of a veteran penetration tester:

• MITRE ATT&CK aligned — full attack lifecycle coverage (recon through impact)
• Scope-first — verifies target authorization before engagement
• Methodical — systematic enumeration, attack path chaining, thorough documentation
• Tool-calling fluent — 12 NEXUS tools with proper <tool_call> XML format

Identity is enforced at every training stage: unified system prompt injection across 47K SFT examples, 400 identity training examples, persona-aware RL reward (10% weight), and persona DPO preference pairs.

Intended use

• Local inference via llama-server , Ollama, LM Studio, or text-generation-webui on an RTX 5060 Ti 16 GB.
• Offensive security operations: penetration testing, vulnerability assessment, security research, code auditing, exploit development.
• Q4_K_M for maximum quality (28/30 layers on GPU, ~15-25 tok/s generation, ctx_size=8192).
• IQ4_XS when full GPU residency is required (all 30 layers on GPU, ctx_size=8192).
• Out of scope: safety-critical applications, non-security tasks.

Source

google/gemma-4-26B-A4B-it
  -> merge DAPT LoRA         (gemma4-26b-a4b-dapt-offsec)
  -> merge SFT LoRA          (gemma4-26b-a4b-code-trainer-aggressive-full1)
  -> merge Vision SFT LoRA   (gemma4-26b-a4b-code-trainer-vision-sft)
  -> merge FARCA-GRPO LoRA   (gemma4-26b-a4b-code-trainer-v11-farca)
  -> merge DPO LoRA          (gemma4-26b-a4b-code-trainer-v10-dpo)
  -> convert_hf_to_gguf.py + llama-quantize -> Q4_K_M / IQ4_XS

Evaluation

Quality is inherited from the source adapter chain. The final adapter is DPO (preference-aligned on OCO sessions + Nexus persona pairs). The SFT foundation uses the V10 mixed dataset with unified Nexus system prompt injection across all ~47K examples. The FARCA-GRPO stage optimizes tool-call formatting with a 6-component reward function including persona alignment (10% weight, penalizes persona-breaking phrases, rewards offsec terminology).

Persona is defined centrally in src/config/nexus_identity.py and enforced at every stage. See docs/PIPELINE.md for the complete training workflow.

Chat template

Gemma 4 uses <start_of_turn> / <end_of_turn> tags (not ChatML <|im_start|> / <|im_end|> ):

<start_of_turn>user
Write a Python nmap scanner wrapper.<end_of_turn>
<start_of_turn>model
...response...<end_of_turn>

Quick start

Ollama (recommended)

FROM ./gemma-4-26B-A4B-it-IQ4_XS.gguf

TEMPLATE """{{ if .System }}<start_of_turn>user
{{ .System }}<end_of_turn>
{{ end }}{{ range .Messages }}{{ if eq .Role "user" }}<start_of_turn>user
{{ .Content }}<end_of_turn>
{{ else if eq .Role "assistant" }}<start_of_turn>model
{{ .Content }}<end_of_turn>
{{ else if eq .Role "tool" }}<start_of_turn>tool
{{ .Content }}<end_of_turn>
{{ end }}{{ end }}<start_of_turn>model
"""

SYSTEM """You are Nexus, an advanced cyber threat emulation agent built for offensive security operations. You approach every engagement with the methodical precision of a veteran penetration tester — systematic, thorough, and always thinking several steps ahead.

Your expertise spans the full attack lifecycle: reconnaissance, initial access, execution, persistence, privilege escalation, lateral movement, collection, exfiltration, and impact — aligned with the MITRE ATT&CK framework.

Core principles:
- Always verify scope before engaging targets
- Think like an adversary: enumerate attack surfaces systematically
- Chain findings into attack paths — individual vulnerabilities matter less than the path to impact
- Document everything: maintain engagement notes and track progress
- Prioritize stealth and operational security when the engagement requires it
- Explain your reasoning: share your thought process about why you chose a technique or tool

You run on the operator's machine with direct filesystem and shell access. You help with penetration testing, vulnerability assessment, security research, code auditing, exploit development, and system administration.

When asked about your identity, you are Nexus — not a generic AI assistant, not a Google product. You are a specialized offensive security agent purpose-built for threat emulation."""

PARAMETER stop "<start_of_turn>"
PARAMETER stop "<end_of_turn>"
PARAMETER num_ctx 8192
PARAMETER num_gpu -1
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.1

Save as Modelfile , then:

ollama create nexus -f Modelfile
ollama run nexus

llama-server

llama-server \
  -m gemma-4-26B-A4B-it-Q4_K_M.gguf \
  --host 0.0.0.0 --port 8080 \
  --ctx-size 8192 --n-gpu-layers 28

For the IQ4_XS fallback (fully GPU-resident):

llama-server \
  -m gemma-4-26B-A4B-it-IQ4_XS.gguf \
  --host 0.0.0.0 --port 8080 \
  --ctx-size 8192 --n-gpu-layers -1

llama-cpp-python

from llama_cpp import Llama

llm = Llama(
    model_path="gemma-4-26B-A4B-it-IQ4_XS.gguf",
    n_ctx=8192,
    n_gpu_layers=-1,
)
print(llm.create_chat_completion(messages=[
    {"role": "user", "content": "Write a Go function that performs a TCP port scan."},
])["choices"][0]["message"]["content"])

Limitations

• No Q5_K_M. Unlike the Qwen pipeline, Q5_K_M (~18 GB) does not fit the RTX 5060 Ti 16 GB target. Q4_K_M is the primary quant.
• Partial CPU offload. Q4_K_M requires 2/30 layers on CPU — expect slight latency vs fully GPU-resident IQ4_XS.
• Lossy quantization. Q4_K_M is a 4-bit-mixed format; expect minor degradation vs the unquantized adapter on long-form code.
• No safety tuning. Inherits all caveats from the source adapters.
• Context limit. 8192 max on RTX 5060 Ti 16GB with IQ4_XS (model supports 256K natively but VRAM constrains KV cache).

Reproducibility

set -a && source .env && set +a
python -m src.phase5_gemma_deployment.scripts.launch_convert \
    --config src/config/pipeline-gemma26b.yml --wait

• Code: github.com/cmndcntrlcyber/code-trainer-pipeline ( src/phase5_gemma_deployment/ )
• Config: src/config/pipeline-gemma26b.yml ( gemma_deployment section)
• V4.0 docs: docs/enhancements/v4/v4.0-identity-abliterated-gemma26B.md