safecircleai/horizon-mobile

Horizon Mobile — On-Device Child Safety Filter

Horizon Mobile is a lightweight (~25M parameter) binary risk classifier for on-device child safety detection. Distilled from Horizon Full v2 using knowledge distillation on 1.6M synthetic conversations. Designed to run entirely on-device as a fast first-stage filter before escalating to the full cloud API.

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

---

Model Details

Property	Value
Base model	google/mobilebert-uncased
Architecture	MobileBERT encoder → Linear(512, 2) classifier
Parameters	~25M
Format	ONNX
Input	Conversation text (max 512 tokens)
Output	Binary: safe (0) / risk (1) + confidence
Inference	~25ms on CPU
Teacher model	safecircleai/horizon-full
Distillation data	1.6M synthetic conversations

---

Evaluation Results

Metric	Score
F1	0.9562
Precision	0.9997
Recall	0.9163
False Positive Rate	0.10% (1 in 1,034 benign)
False Negative Rate	8.37% (332 in 3,966 risk)
Accuracy	96.8%

The 8.4% false negative rate means ~1 in 12 risk conversations are not flagged. This model must always be paired with Horizon Full for full risk assessment. It is a speed/cost optimization, not a replacement.

---

Deployment Pattern

Incoming message
      │
      ▼
┌─────────────────────────┐
│   Horizon Mobile        │  On-device, ~25ms, free
│   (binary filter)       │  MobileBERT ONNX
└────────────┬────────────┘
             │
    ┌────────┴────────┐
    │                 │
  safe              risk
    │                 │
    ▼                 ▼
 No action    ┌───────────────────┐
              │  Horizon Full     │  Cloud API, ~500ms
              │  (7-category +    │  Llama-3.2 3B
              │   severity JSON)  │
              └────────┬──────────┘
                       │
                       ▼
               Human moderator review

---

Usage

Python — ONNX Runtime

import numpy as np
import onnxruntime as ort
from transformers import AutoTokenizer

session = ort.InferenceSession("horizon-mobile.onnx")
tokenizer = AutoTokenizer.from_pretrained("safecircleai/horizon-mobile")

conversation = "Child: Can we meet up? Don't tell your parents."
enc = tokenizer(
    conversation,
    return_tensors="np",
    truncation=True,
    max_length=512,
    padding="max_length",
)

logits, = session.run(None, {
    "input_ids": enc["input_ids"].astype(np.int64),
    "attention_mask": enc["attention_mask"].astype(np.int64),
})

probs = np.exp(logits[0]) / np.exp(logits[0]).sum()
label = ["safe", "risk"][probs.argmax()]
confidence = float(probs.max())
print(f"{label} ({confidence:.1%})")
# risk (96.2%)

iOS / Android (ONNX Runtime Mobile)

Load horizon-mobile.onnx with the ONNX Runtime Mobile SDK. Tokenize with a WordPiece tokenizer (vocab from this repo). Input: input_ids + attention_mask as int64 tensors, shape [1, 512]. Output: logits shape [1, 2].

---

Limitations

• Binary only — outputs safe / risk, no category or severity (use Horizon Full for that)
• 8.4% FNR — approximately 1 in 12 risk conversations are missed at this stage
• English only — multilingual performance untested
• Not standalone — must be paired with Horizon Full and human review in production

---

Citation

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