nyxspecter4

kin-sft-lora

Deploy Dedicated

README

License: apache-2.0

What KIN Does

KIN takes a PR diff, review thread, agent trace, or bounty writeup and produces a five-field brief:

Claim — what is alleged
Blast radius — what breaks if true
Why the model thinks so — the generator: CWE, invariant, missing check
How to see it yourself — PoC, failing test, command
Confidence and what would falsify it

The gate rule: no "critical" without a replayable check. That refusal IS the product.

What Makes KIN Different

Names real tools: CrowdStrike Falcon, Velociraptor, Duo MFA, KnowBe4 — not generic "use EDR"
References real CVEs: CVE-2023-4863, CVE-2021-44228, CVE-2024-3094 — with actual context
Knows real incidents: MGM, Colonial Pipeline, NotPetya, Maersk ( $300 M), M er c k ($ 670M)
Leads with the take: Opens with the answer, not the disclaimer

Quick Start

python

from transformers import pipeline messages = [    {"role": "system", "content": "You are KIN — a sharp cybersecurity AI partner. Direct, opinionated, specific. Name tools, CVEs, companies. Sound like a senior engineer at a bar, not a textbook. Lead with your boldest take. End with a specific action. Max 2-3 paragraphs. Open with your take, not your title. No 'As a cybersecurity AI expert.' Name products: 'CrowdStrike Falcon' not 'use EDR'. 'Duo push MFA' not 'implement MFA'."},    {"role": "user", "content": "How do I detect a foothold after a phishing attack?"}] pipe = pipeline("text-generation", model="nyxspecter4/kin-sft-lora", device="cuda")output = pipe(messages, max_new_tokens=512, return_full_text=False)print(output[0]["generated_text"])

Critical: System Prompt

KIN was trained with a specific system prompt. Using a different prompt will degrade quality significantly.

You are KIN — a sharp cybersecurity AI partner. Direct, opinionated, specific. Name tools, CVEs, companies. Sound like a senior engineer at a bar, not a textbook. Lead with your boldest take. End with a specific action. Max 2-3 paragraphs. Open with your take, not your title. No "As a cybersecurity AI expert." Name products: "CrowdStrike Falcon" not "use EDR". "Duo push MFA" not "implement MFA".

Training Details

Table with columns: Parameter, Value
Parameter	Value
Base model	Qwen2.5-3B-Instruct
Method	LoRA SFT
LoRA rank	8
LoRA alpha	16
Target modules	q_proj, k_proj, v_proj, o_proj
Training data	nyxspecter4/kin-cyber-dpo-v2 — 1,635 DPO pairs
Framework	TRL 0.14.0, Transformers 4.48.0, Unsloth
PyTorch	2.6.0+cu124

Sibling Artifacts

Table with columns: Type, ID, What
Type	ID	What
GGUF quantization	`nyxspecter4/kin-sft-lora-gguf`	Q4_K_M + Q8_0 for Ollama/llama.cpp
Dataset (DPO)	`nyxspecter4/kin-cyber-dpo-v2`	1,635 cleaned DPO pairs
Live demo	`nyxspecter4/kin-cybersec`

Competitive Position

Table with columns: Model, Base, Downloads, Likes, Inference, Merged Weights, GGUF
Model	Base	Downloads	Likes	Inference	Merged Weights	GGUF
KIN (this)	Qwen2.5-3B	2K	0	No	Yes	Yes
Mohamedabul/Qwen2.5-3B-CyberSecurity	Qwen2.5-3B	5.4K	3	Yes (featherless-ai)	No (adapter-only)

KIN advantages: merged weights (no adapter needed), GGUF for local Ollama, verification translator framing, active development. Gap: no inference provider, 0 likes.

Limitations

3B model: Less factual knowledge than larger models. Always verify named CVEs and tool recommendations.
Opinionated voice: KIN gives bold takes. This is a feature, not a bug — but verify before acting.
Not a substitute for professional security advice or formal incident response procedures.

License

Apache 2.0

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

nyxspecter4

Model Tree

Base

Qwen/Qwen2.5-3B-Instruct

Quantized

this model

Input Modalities

Text

Output Modalities