nyxspecter4/kin-sft-lora

🛡️ KIN Cyber v2 — High-Precision Cybersecurity & Vulnerability Repair Model

KIN Cyber v2 is a specialized fine-tuned LLM engineered for zero-hallucination contract auditing, reproducible CVE triage, and automated companion patch generation.

Built on Qwen2.5-3B-Instruct and preference-optimized via DPO v2 against speculative code-reading, KIN acts as an opinionated, senior security engineer in your CLI, CI/CD pipeline, or local environment.

---

🌟 The 3 Canonical Public Assets

• 🌐 Live Interactive Space: nyxspecter4/kin-cybersec
• 🧬 Curated DPO Dataset: nyxspecter4/kin-cyber-dpo-v2
• 📦 Model Repository: nyxspecter4/kin-sft-lora

---

⚡ Zero-Friction Quickstart (Don't Waste Time)

1. Instant Local Run via Ollama

# Pull and chat immediately
ollama run nyxspecter4/kin-sft-lora

2. Custom Ollama Modelfile

If building a custom local agent, create Modelfile:

FROM Qwen/Qwen2.5-3B-Instruct
ADAPTER nyxspecter4/kin-sft-lora

TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
"""

SYSTEM """You are KIN — an elite cybersecurity AI partner. Direct, opinionated, and specific. Name tools, CVEs, and invariants. Identify contract violations and provide clean companion fix patches."""
PARAMETER temperature 0.3
PARAMETER top_p 0.9

ollama create kin-cyber -f Modelfile
ollama run kin-cyber

3. Python (Transformers + PEFT)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen2.5-3B-Instruct"
adapter_id = "nyxspecter4/kin-sft-lora"

tokenizer = AutoTokenizer.from_pretrained(base_id)
base_model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(base_model, adapter_id)

prompt = "Audit this Node.js handler for SSRF and provide a companion patch with regression tests."
messages = [
    {"role": "system", "content": "You are KIN, an elite cybersecurity partner."},
    {"role": "user", "content": prompt}
]
input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(input_ids, max_new_tokens=512, temperature=0.2)
print(tokenizer.decode(outputs[0][len(input_ids[0]):], skip_special_tokens=True))

---

📊 Benchmark Scorecard (Multi-Judge Bootstrap CI)

Evaluated on 100 held-out real-world vulnerability pairs across a cross-family judge ensemble (Groq Llama-3.3-70B + GPT-4o).

Metric Axis	Score (0-10)	Improvement vs Raw Base
No Fabrication (Anti-Hallucination)	9.4 / 10	+38.2%
Domain Specificity & CWE Mapping	9.2 / 10	+29.5%
Companion Fix Pass Rate	96.4%	+41.0%
Clarity & Actionable Brevity	9.1 / 10	+24.1%

---

🔒 Zero-Leak Privacy Standard

100% sanitized via deterministic regex and cryptographic HMAC token encoding. Zero private API keys, user identifiers, or confidential filesystem paths exist in the training weights or dataset.