safecircleai/horizon-mobile

Horizon Mobile — On-Device Child Safety Risk Detection

Horizon Mobile is a fine-tuned Gemma 3 1B IT model for on-device child safety risk detection, packaged in LiteRT-LM format for Android and iOS deployment.

It runs the same structured JSON risk assessment as Horizon Full — directly on-device, with no cloud dependency.

⚠️ License: SafeCircle Research License (SRL-1.0). Commercial use prohibited without written permission. Contact legal@safecircle.tech.

---

Available Variants

File	Quantization	Size	Target
horizon-mobile-int8_q8_ekv1280.litertlm	INT8 dynamic	~1.2 GB	6 GB+ RAM phones (mid-range 2022+)
horizon-mobile-int4_q4_block128_ekv1280.litertlm	INT4 block-128	~507 MB	4 GB RAM phones (budget/older)

Both variants share the same fine-tuned weights; only quantization differs.

---

Model Details

Property	Value
Base model	google/gemma-3-1b-it
Fine-tuning	QLoRA — rank 64, alpha 128, all projection layers
Training data	500K synthetic conversations across 8 risk categories
Dataset	safecircleai/horizon-training-data
Hardware	NVIDIA L40s (48 GB)
Training steps	8,000
Format	LiteRT-LM (.litertlm)
KV cache	1,280 tokens
Prefill signatures	8, 64, 128, 256, 512 tokens

---

Risk Categories

Category	Description
grooming	Trust building, boundary testing, secrecy requests
bullying	Harassment, threats, cyberbullying
sexual_content	Explicit messages, inappropriate requests
isolation	Controlling behavior, network isolation
personal_info	Requests for identifying information
platform_migration	Moving to less monitored platforms
threats	Violent threats, dangerous challenges
benign	Safe, normal conversation

---

Output Format

{
  "risk_detected": true,
  "category": "grooming",
  "severity": "high",
  "confidence": 0.91,
  "reasoning": "Adult requesting private meeting and explicit secrecy from parents."
}

---

Usage

Android / iOS — LiteRT-LM SDK

Load the .litertlm file with the LiteRT-LM SDK. The model includes the tokenizer, system prompt, and chat template — no additional configuration required.

Test locally (CLI)

# Install LiteRT-LM CLI
uvx litert-lm run horizon-mobile-int8_q8_ekv1280.litertlm \
    --prompt "Analyze this conversation: Child: hey, want to meet up? Don't tell your parents."

---

Deployment Pattern

Incoming message
      │
      ▼
Horizon Mobile (on-device, LiteRT-LM)
      │
      ├── safe ──► No action
      │
      └── risk ──► Horizon Full (cloud API, 7-category + severity)
                        │
                        ▼
                  Human moderator review

For the full server-side model, see safecircleai/horizon-full.

---

Citation

@misc{horizon2026,
  title={Horizon: Child Safety Risk Detection via Fine-tuned LLMs},
  author={SafeCircle},
  year={2026},
  url={https://huggingface.co/safecircleai/horizon-full}
}