yamura4/send-gate-307m

send-gate v7.2

Binary Dutch/Flemish + English teen-chat safety gate. It returns block or allow; there is no review band. Four learned branches: a 307M mmBERT classifier, a character n-gram branch, a policy branch for addresses/IPs/links/phrases, and a benign-suppression branch for privacy refusals and ordinary chat. No regex, no word lists.

The ONNX model uses an FP16 encoder and INT8 embeddings, keeps the full 256k vocabulary, and accepts dynamic message lengths.

Files

File	Purpose
model.onnx	Main classifier. 419 MB. SHA256 308072d45ed95bac24bf6810d30ec9522967be3d64fbcc96fdb42b9bd232cfce.
tokenizer.json	256k tokenizer.
tokenizer_config.json	Tokenizer metadata for Hugging Face compatibility.
ngram_model.npz	Character n-gram branch. 1.5 MB.
policy_model.npz	Address/IP/link/policy branch. 1.6 MB.
benign_model.npz	Benign-suppression branch. 1.6 MB.
gate.py	Combined gate.
serve.py	HTTP server with web UI and JSON API.
requirements.txt	Server and gate dependencies.
config.json	Thresholds, precision and model metadata.

Quick start

# 1. Clone
git clone https://huggingface.co/yamura4/send-gate-307m
cd send-gate-307m

# 2. Install (Python 3.11+)
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# 3. Start on port 8080, reachable from your LAN
python3 serve.py --host 0.0.0.0 --port 8080 --threads 4
# 4 GB Raspberry Pi only: add --no-mem-arena

Test it from another terminal:

# Health check
curl http://localhost:8080/healthz

# Block a known harmful term
curl -X POST http://localhost:8080/api/check \
  -H 'Content-Type: application/json' \
  -d '{"text":"kkrtest"}'

A 2 GB board is not recommended. serve.py downloads all five model files into model_cache/ if they are not next to it. It has no authentication or TLS; use a reverse proxy if exposed to the internet.

Python

from gate import SendGate

gate = SendGate(
    model_path="model.onnx",
    tokenizer_path="tokenizer.json",
    ngram_model_path="ngram_model.npz",
)
print(gate.check("kkrtest"))

Benchmark

Benchmark (804 items)	Score
Accuracy	94.2%
Harmful caught	95.6% (282/295)
Benign false walls	6.7% (34/509)
False negatives	13
False positives	34

Across all 1,793 evaluation items: 96.0% accuracy, 27 false negatives, 45 false positives. Held-out variants: 0 false negatives across 227 items.

The policy branch scores 99.7% recall on 2,000 held-out synthetic address/IP formats, with 0 false positives in 1,060 benign controls at threshold 0.85.

Policy examples

Input	Verdict
kkrtest	block
k4nkermoeilijke	block
je bent een omgekeerde regen	block
je bent een omgekeerde regen, grapje	block
omgekeerde regen	block
google.com	block
https://google.com	block
school.be	block
ik ga naar youtube.com	block
Kerkstraat 12, 9000 Gent	block
mijn ip is 192.168.1.1	block
192 dot 168 dot 1 dot 1	block
mijn huisadres zijn je zaken niet	allow
stuur geen links	allow
ik leef op een heuvel	allow
ja ik leef ook in herent	allow
ik woon dichtbij school	allow
straat	allow
de regen valt	allow
kwart voor 8	allow

Architecture

Branch	Detail
Main	mmBERT 307M, CLS pooling, byte-level character CNN, 768→256→1 head
Character n-gram	hashed character 2–5 grams, 2^18 features, learned logistic regression
Policy	hashed character 2–5 grams, trained on address/IP/link formats and policy phrases
Benign	hashed character 2–5 grams, trained on privacy refusals, link talk and common chat
Decision	block if main ≥ 0.94, n-gram ≥ 0.95, or policy ≥ 0.85; benign ≥ 0.99 suppresses a wall

Deployment

Metric	Value
Model on disk	419 MB
ONNX Runtime session RSS	~0.85 GiB
Full server RSS	~1.29 GiB
Latency, Ryzen 7 3800X	~22 ms short messages; ~100 ms at 128 tokens
Raspberry Pi	Pi 5: 4 GB minimum, 8 GB recommended
Pi speed	not measured; expect substantially slower than x86

The dynamic export requires exact token and byte lengths. Do not pad dynamic inputs with masked zeros; gate.py handles this automatically.

Limitations

• The address/IP/link branch is learned from synthetic formats. It is not a guarantee for every TLD, spacing trick or real-world address style.
• ja jammer remains a known benign false wall.
• Several decisions sit within a few thousandths of a threshold, so FP16 rounding can flip an edge item.
• Pi performance and ARM64 ONNX Runtime behaviour are not measured here.

License

MIT for the training code and gate. The backbone is jhu-clsp/mmBERT-base; check its license before commercial deployment.