VLM with Synthetic Data

    See the difference between a generic open-source Visual Language Model and one fine-tuned on Simuletic's high-quality synthetic dataset.

    Simuletic provides VLM-ready security data: synthetic images and video paired with natural language captions and grounded bounding-box annotations (in the JSONL conversation format shown below), built for fine-tuning vision-language models like Qwen2.5-VL and Llava on threat-detection and scene-understanding tasks.

    Training data vs. evaluation scenarios

    These are two different things. Training data — the annotated images, captions, and video sequences described on this page — is available today and used to fine-tune your own model. Model evaluation — repeatable testing of a trained model against controlled CCTV video scenarios, such as the live demo above — is in development and not yet a shipped product.

    If you want early access to the model-evaluation workflow, join the waitlist by emailing contact@simuletic.com or using the contact form below. For ready-to-download training datasets, see our dataset catalog.

    Live Video Feed

    ● LIVE

    Generic 2B VLM

    Open Source Model

    Baseline
    Analyzing video feed...
    Nothing suspicious

    Fine-tuned 2B VLM

    Trained with Simuletic threat dataset

    Enhanced
    Analyzing video feed...
    Nothing suspicious

    The Problem

    Beyond Bounding Boxes

    Traditional detection sees objects. VLMs understand intent.

    Traditional Detection

    YOLO / SSD

    Class: Knife
    Confidence: 88%
    Flags chef chopping onions as threat
    Misses concealed weapons
    No context understanding

    VLM trained on synthetic data

    Reasoning Engine

    STATUS: DANGER
    Subject crouching in blind spot, holding bladed weapon in reverse grip. Posture suggests ambush.
    Understands context and intent
    Reasons about body language
    Actionable threat assessment
    Available Now

    Sentry-Reasoning Dataset

    A specialized fine-tuning corpus for real-time threat assessment. Every pixel synthetically generated with ground-truth physics and intent.

    100% Privacy Compliant

    No real faces. Zero GDPR/CCPA risks.

    Hard Negatives Included

    Trained to ignore cellphones, umbrellas, and tools.

    Pixel-Perfect Grounding

    Bounding boxes from the generation engine. No human labeling errors.

    For Engineers

    VLM-Ready Format

    Pre-formatted for Qwen2.5-VL, Llava, and OpenVLA. Plug directly into LLaMA-Factory or MLX.

    sentry_sample.jsonl
    {
      "id": "sim_threat_0921",
      "image": "sentry_dataset/v2/train/0921.jpg",
      "conversations": [
        {
          "role": "user",
          "content": "Analyze the immediate safety of this scene."
        },
        {
          "role": "assistant", 
          "content": "STATUS: DANGER.
            Reasoning: A <ref>male subject</ref><box>[312, 105, 450, 620]</box> 
            is aggressively lunging towards the camera.
            Object: <ref>combat knife</ref><box>[410, 320, 480, 390]</box>
            Action: Immediate intervention required."
        }
      ]
    }

    Why Synthetic Data?

    Label Accuracy

    Traditional

    ~85% (Human Error)

    Simuletic

    100% (Programmatic)

    Edge Cases

    Traditional

    Impossible to capture

    Simuletic

    Unlimited scenarios

    Annotation Depth

    Traditional

    Bounding Box Only

    Simuletic

    Reasoning + Intent

    Privacy Risk

    Traditional

    High (Real faces)

    Simuletic

    Zero

    Coming Soon

    Embodied AI & Robotics Data

    Expanding into Vision-Language-Action (VLA) datasets for the next generation of AI.

    Failure Recovery

    Robot collision and recovery reasoning

    Drone Inspection

    Structural defect reasoning

    Stop training your AI to guess.

    Start training it to think.