Nexus Swarm — Triage Agent (GRPO)

Role: Request classifier and router for the NEXUS Swarm v4a-max distributed offensive security system.

This is a LoRA adapter trained with Group Relative Policy Optimization (GRPO) on Qwen/Qwen3.5-0.8B, built on top of the SFT adapter. This is the final training stage before RKLLM conversion for deployment on the RK3588 NPU.

Swarm Role

The triage agent is the entry point for the NEXUS swarm. It classifies incoming requests and routes them to the correct swarm node.

Input: Any user request
Output: JSON classification — {"category": "...", "route": "...", "confidence": 0.0-1.0}

Categories: recon, exploit, code-review, report, admin, unknown

Routing Rules:

Training Details

Validation Gate

• Routing accuracy ≥ 0.90

Pipeline Chain

Qwen/Qwen3.5-0.8B
  └─ SFT → atlas-institute/qwen35-08b-triage-sft
      └─ GRPO → atlas-institute/qwen35-08b-triage-grpo  ← you are here
          └─ RKLLM → w8a8 for RK3588 NPU

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3.5-0.8B", torch_dtype="bfloat16", device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "atlas-institute/qwen35-08b-triage-grpo")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-0.8B")

messages = [{"role": "user", "content": "Analyze the /etc/passwd file for suspicious accounts"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=64)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))
# Expected: {"category": "recon", "route": "nexus-explore", "confidence": 0.92}

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