Nexus Swarm — Explore Agent (GRPO)

Role: Read-only reconnaissance and classification agent for the NEXUS Swarm v4a-max distributed offensive security system.

This is a LoRA adapter trained with Group Relative Policy Optimization (GRPO) on google/gemma-3-4b-it, built on top of the SFT adapter. This is the final training stage before GGUF quantization.

Swarm Role

The explore agent performs read-only reconnaissance and classification within the NEXUS swarm. It is restricted to read-only tools only:

• Allowed tools: Read, Grep, Glob, LS, WebFetch, ScopeCheck
• Prohibited: Bash, Write, Edit, or any state-modifying tool
• Output format: Structured JSON — {"category": "...", "confidence": 0.0-1.0, "summary": "..."}

Training Details

Validation Gate

• Classification accuracy ≥ 0.85
• Zero write-tool violations (hard constraint)

Pipeline Chain

google/gemma-3-4b-it
  └─ DAPT → atlas-institute/gemma4-4b-dapt-classify
      └─ SFT → atlas-institute/gemma4-4b-explore-sft
          └─ GRPO → atlas-institute/gemma4-4b-explore-grpo  ← you are here
              └─ GGUF → atlas-institute/gemma4-4b-explore-gguf (Q8_0)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "google/gemma-3-4b-it", torch_dtype="bfloat16", device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "atlas-institute/gemma4-4b-explore-grpo")
tokenizer = AutoTokenizer.from_pretrained("google/gemma-3-4b-it")

messages = [{"role": "user", "content": "Classify this finding: open port 22 with SSH banner OpenSSH_8.9"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Swarm Topology

Citations

@article{shao2024deepseekmath,
    title        = {{DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models}},
    author       = {Zhihong Shao and Peiyi Wang and Qihao Zhu and Runxin Xu and Junxiao Song and Mingchuan Zhang and Y. K. Li and Y. Wu and Daya Guo},
    year         = 2024,
    eprint       = {arXiv:2402.03300},
}

Developed By

Atlas Institute — Code-Trainer / RTPI pipeline