Nexus Swarm Orchestrator — SFT Adapter (Qwen3-14B)

LoRA adapter fine-tuned on Qwen/Qwen3-14B for the orchestrator role in the NEXUS Swarm — a distributed multi-model offensive security agent system.

This is the SFT stage of the orchestrator training pipeline. The full chain is:

DAPT → SFT → FARCA-GRPO → DPO → GGUF

Role: Orchestrator

The orchestrator is the central reasoning node of the Nexus swarm. It runs on an RTX 5060 Ti (16 GB) and is responsible for:

• Decomposing complex offensive security tasks into subtasks
• Delegating tool-heavy or scan-parsing work to the worker node
• Delegating read-only search and classification to the explore agent
• Synthesizing results from delegated tasks into coherent responses
• Following a structured ASSESS → SELECT → ACT reasoning protocol before each tool call

Training Details

Validation Gate

The model must pass these thresholds before proceeding to RL:

• Tool-call pass rate ≥ 0.75 across 14 scenarios
• GSM8K accuracy ≥ 0.55
• Agent progress ≥ 3 across 7 agentic scenarios

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-14B", torch_dtype="auto", device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "atlas-institute/qwen3-14b-orchestrator-sft")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")

messages = [
    {"role": "system", "content": "You are Nexus, an offensive security orchestrator agent."},
    {"role": "user", "content": "Enumerate services on 10.10.10.5 and identify potential attack vectors."},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=768)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Related Models

Swarm Topology

Developed by

Atlas Institute · Part of the Code-Trainer (RTPI) project.