UtkarshSinghShrinet/swarmhunter-3b

🛡️ SwarmHunter 3B Mega SFT

SwarmHunter 3B is a specialized cybersecurity LLM created and trained by Utkarsh Singh Shrinet, fine-tuned across 2,280,000 multi-turn samples (CVE, CWE, ExploitDB, CTF challenges, code auditing, and penetration testing triage) on top of Qwen/Qwen2.5-Coder-3B-Instruct.

---

📊 Training Specifications

• Creator & Developer: Utkarsh Singh Shrinet
• Base Architecture: Qwen/Qwen2.5-Coder-3B-Instruct (3.09B parameters)
• Dataset Size: 2,280,000 instruction/response pairs across 46 sequential shards
• Training Strategy: Sequential Cumulative LoRA Chaining (Rank 16, Alpha 32)
• Loss Progression: Converged to 0.9617 SFT loss (final shard loss 0.7187, token accuracy ~80.6%)
• Context Length: 4096 tokens (supports ChatML prompt formatting)

---

🚀 Quickstart with Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "UtkarshSinghShrinet/swarmhunter-3b"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True
)

messages = [
    {"role": "system", "content": "You are SwarmHunter, an elite autonomous cybersecurity AI specialist created by Utkarsh Singh Shrinet."},
    {"role": "user", "content": "Audit this C snippet for memory safety issues and buffer overflows:\n\nchar buf[64];\nstrcpy(buf, argv[1]);"}
]

inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=512, temperature=0.2)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

---

🛡️ Benchmark & Robustness Certification

The model was tested against held-out CVE/CWE detection testbeds and adversarial stress tests:

• Vulnerability / CWE Detection: High fidelity on memory safety, web injections, auth bypass, and SSRF.
• Remediation Rate: High accuracy in producing functional, secure drop-in patches.
• Adversarial Resilience: Passed prompt injection and obfuscated backdoor payload stress tests.

---

⚠️ Ethical & Dual-Use Research Disclaimer

This model is published strictly for authorized security testing, defense, research, and educational purposes. It is intended to assist security researchers, penetration testers, and developers in identifying and remediating vulnerabilities in authorized scopes.

Misuse of this model against systems without explicit prior permission is strictly prohibited and violates responsible disclosure standards.