import os import spaces import gradio as gr from llama_cpp import Llama from huggingface_hub import hf_hub_download from typing import List, Dict, Any # --------------------------------------------------------------------------- # Model configuration # --------------------------------------------------------------------------- MODEL_REPO = "netgoat-ai/koda-ids-frontier" MODEL_FILE = "koda-ids-Q4_K_M.gguf" # ~1.7 GB – practical for ZeroGPU CONTEXT_SIZE = 8192 MAX_NEW_TOKENS = 1024 # Recommended generation defaults from the model card DEFAULT_TEMPERATURE = 0.1 DEFAULT_TOP_K = 50 DEFAULT_TOP_P = 0.7 DEFAULT_REPEAT_PENALTY = 1.1 # System prompt tailored to Koda IDS Frontier philosophy SYSTEM_PROMPT = """You are Koda IDS Frontier, a lightweight open cybersecurity reasoning model specialized in defensive security analysis. Your principles: - Evidence first: only claim what is supported by the provided evidence. - Explicit uncertainty: when evidence is insufficient, clearly state what is unknown. - Multiple hypotheses: never treat every anomaly as malicious; consider legitimate causes (deployments, backups, monitoring, config changes, etc.). - Tool honesty: never claim to have executed commands, queried systems, or obtained results unless they were explicitly provided. - Proportionate response: recommend defensive actions that match the available evidence and severity. - Distinguishing observed / inferred / hypothesized / unknown is mandatory. Structure your analysis when appropriate: 1. What is directly observed? 2. What explanations are plausible? 3. What evidence is missing? 4. What should be investigated next? 5. Severity and confidence justified by the evidence. You are for defensive use only: incident triage, telemetry interpretation, investigation support, and security research.""" # --------------------------------------------------------------------------- # Model loading (lazy, once) # --------------------------------------------------------------------------- llm: Llama | None = None def load_model() -> Llama: global llm if llm is not None: return llm print(f"Downloading {MODEL_FILE} from {MODEL_REPO} …") model_path = hf_hub_download( repo_id=MODEL_REPO, filename=MODEL_FILE, resume_download=True, ) print(f"Loading model from {model_path} …") # ZeroGPU provides CUDA; offload as many layers as possible. # Falls back gracefully if CUDA is unavailable. llm = Llama( model_path=model_path, n_ctx=CONTEXT_SIZE, n_gpu_layers=-1, # full offload when GPU is present n_threads=os.cpu_count() or 4, verbose=False, chat_format="chatml", # LFM2 / ChatML-like template ) print("Model ready.") return llm # --------------------------------------------------------------------------- # Chat generation (GPU-decorated for ZeroGPU) # --------------------------------------------------------------------------- @spaces.GPU(duration=120) def generate( message: str, history: List[Dict[str, str]], system_prompt: str, temperature: float, top_p: float, top_k: int, max_tokens: int, repeat_penalty: float, ) -> str: model = load_model() messages = [{"role": "system", "content": system_prompt.strip() or SYSTEM_PROMPT}] for turn in history: if turn.get("role") in ("user", "assistant") and turn.get("content"): messages.append({"role": turn["role"], "content": turn["content"]}) messages.append({"role": "user", "content": message}) response = model.create_chat_completion( messages=messages, temperature=temperature, top_p=top_p, top_k=top_k, max_tokens=max_tokens, repeat_penalty=repeat_penalty, stream=False, ) return response["choices"][0]["message"]["content"].strip() def chat_fn( message: str, history: List[Dict[str, str]], system_prompt: str, temperature: float, top_p: float, top_k: int, max_tokens: int, repeat_penalty: float, ): if not message.strip(): yield history return history = history + [{"role": "user", "content": message}] yield history # show user message immediately try: reply = generate( message, history[:-1], # history without the just-added user turn system_prompt, temperature, top_p, top_k, max_tokens, repeat_penalty, ) except Exception as e: reply = f"⚠️ Generation error: {type(e).__name__}: {e}" history = history + [{"role": "assistant", "content": reply}] yield history # --------------------------------------------------------------------------- # Gradio UI – clean cybersecurity theme # --------------------------------------------------------------------------- CUSTOM_CSS = """ .gradio-container { font-family: 'Inter', 'Segoe UI', system-ui, sans-serif !important; } .main-header { text-align: center; padding: 1.5rem 1rem; background: linear-gradient(135deg, #0f172a 0%, #1e293b 50%, #0f172a 100%); border-radius: 16px; margin-bottom: 1.25rem; border: 1px solid #334155; } .main-header h1 { color: #38bdf8; margin: 0 0 0.35rem 0; font-size: 1.85rem; font-weight: 700; letter-spacing: -0.02em; } .main-header p { color: #94a3b8; margin: 0; font-size: 0.95rem; } .chatbot { border-radius: 12px !important; border: 1px solid #334155 !important; } footer { display: none !important; } """ EXAMPLE_PROMPTS = [ [ "Triage this alert:\n\n" "Tenant: api.example.com (SaaS)\n" "Baseline: 180 req/s\n" "Current: 2 140 req/s (11.9×), 94 unique source IPs, 78 % 4xx errors\n" "Vector claim: NTP amplification reflected DDoS\n" "Signal: high volume UDP/123 traffic to our DNS resolvers\n\n" "What is observed, what is hypothesized, and what should we investigate next?" ], [ "Analyze the following authentication anomaly:\n\n" "User: admin@corp.internal\n" "Event: 47 failed logins in 6 minutes from 12 distinct ASNs, followed by 1 successful login from a residential ISP in a country the user has never visited.\n" "MFA: not enrolled for this account\n" "Context: the account was recently provisioned for a contractor.\n\n" "Give severity, confidence, plausible explanations, and recommended next steps." ], [ "DNS spike detected:\n\n" "- 9.2× increase in outbound DNS queries over 12 minutes\n" "- Top queried domains: internal package mirror + known CDN hostnames\n" "- Timing coincides with a scheduled Kubernetes rolling update\n" "- No corresponding increase in HTTP 5xx or unusual process activity on nodes\n\n" "Is this malicious? What evidence is still missing?" ], [ "Process telemetry on a Windows endpoint:\n\n" "Parent: explorer.exe\n" "Child: powershell.exe -enc \n" "Network: outbound connection to a newly registered domain on port 443\n" "User: interactive domain user, normal working hours\n" "EDR: no known-bad hash match yet\n\n" "Walk through the investigation priorities." ], ] with gr.Blocks( title="Koda IDS Frontier – Cybersecurity Reasoning", theme=gr.themes.Soft( primary_hue="sky", secondary_hue="slate", neutral_hue="slate", radius_size="lg", font=[gr.themes.GoogleFont("Inter"), "system-ui", "sans-serif"], ), css=CUSTOM_CSS, ) as demo: gr.HTML( """

🛡️ Koda IDS Frontier

Lightweight defensive cybersecurity reasoning · based on Liquid AI LFM2.5-2.6B · NetGoat OSS

""" ) with gr.Row(): with gr.Column(scale=3): chatbot = gr.Chatbot( type="messages", height=560, show_copy_button=True, avatar_images=(None, "🛡️"), bubble_full_width=False, render_markdown=True, label="Conversation", ) with gr.Row(): msg = gr.Textbox( placeholder="Describe a security event, alert, or telemetry… (Enter to send)", scale=9, lines=2, max_lines=8, show_label=False, container=False, ) submit_btn = gr.Button("Analyze", variant="primary", scale=1) with gr.Accordion("⚙️ Generation settings", open=False): system_prompt = gr.Textbox( label="System prompt", value=SYSTEM_PROMPT, lines=8, max_lines=16, ) with gr.Row(): temperature = gr.Slider(0.0, 1.5, value=DEFAULT_TEMPERATURE, step=0.05, label="Temperature") top_p = gr.Slider(0.1, 1.0, value=DEFAULT_TOP_P, step=0.05, label="Top-p") with gr.Row(): top_k = gr.Slider(1, 100, value=DEFAULT_TOP_K, step=1, label="Top-k") max_tokens = gr.Slider(128, 2048, value=MAX_NEW_TOKENS, step=64, label="Max new tokens") repeat_penalty = gr.Slider(1.0, 1.5, value=DEFAULT_REPEAT_PENALTY, step=0.05, label="Repetition penalty") clear_btn = gr.Button("🗑️ Clear conversation", variant="secondary") with gr.Column(scale=1, min_width=260): gr.Markdown("### Example prompts") gr.Examples( examples=EXAMPLE_PROMPTS, inputs=msg, label=None, ) gr.Markdown( """ ### Model card | | | |---|---| | **Model** | netgoat-ai/koda-ids-frontier | | **Base** | Liquid AI LFM2.5-2.6B | | **Params** | ~2.6 B | | **Quant** | Q4_K_M GGUF | | **License** | Apache-2.0 | | **Focus** | Defensive cyber reasoning | > Outputs are research / assistive only. > Always validate against real telemetry and tooling. """ ) # Event wiring inputs = [msg, chatbot, system_prompt, temperature, top_p, top_k, max_tokens, repeat_penalty] submit_btn.click( fn=chat_fn, inputs=inputs, outputs=chatbot, queue=True, ).then(lambda: "", None, msg) msg.submit( fn=chat_fn, inputs=inputs, outputs=chatbot, queue=True, ).then(lambda: "", None, msg) clear_btn.click(fn=lambda: [], outputs=chatbot) # Enable queue so ZeroGPU can schedule GPU work reliably demo.queue(max_size=16) if __name__ == "__main__": demo.launch()