send-gate-307m

send-gate-307m

Checks Dutch, Flemish and English teen chat messages and returns block or allow, catching harmful terms plus shared addresses, IPs and links. For builders of child-safe chat apps.

Made by yamura4 on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: send-gate is a safety gate for teen chat in Dutch, Flemish and English. It reads one message and returns either block or allow, with no middle review state. It combines a 307M-parameter classifier with three smaller learned branches for character patterns, addresses and links, and ordinary chat.

Who it suits: It suits developers of chat apps for teenagers who want to stop messages that share addresses, IPs, links or harmful terms before they are sent.

Industry | Who uses it | How they use it | Example
Education technology | School chat app builders | Screen each message before it is sent | Blocking a message that gives out a home address
Youth platforms | Community moderators | Add an automatic first check to moderation | Stopping a link to an outside site in a teen chat
Games | Game chat developers | Filter harmful terms and contact details | Blocking an IP address typed in plain digits
Small devices | Hobbyists with a Raspberry Pi | Run the gate on a local board | A home chat box for a school project

You can skip it if your users write in other languages, since the card only covers Dutch, Flemish and English.

Where to use it:

• Python 3.11 or newer, through the gate script or the small HTTP server that comes with it
• A Raspberry Pi 5, which the card says needs 4 GB at minimum and recommends 8 GB
• An ordinary x86 computer, where the card measured its speed
• Windows, macOS or Linux, with Python for the download

Use case | Who | What they get back
Check one message | Chat developers | A block or allow verdict
Spot shared addresses and links | Moderators | A block for messages that carry contact details
Run a local check service | Hobbyists | A server with a web page and a JSON interface

How to set it up:

• Python 3.11 or newer
• Disk space, about 419 MB for the main model plus small extra files
• About 1.29 GiB of memory for the full server, per the card

1. Get the repository files to your computer.
2. Make a Python virtual environment and activate it.
3. Install the requirements.
4. Start the server on a port you choose.

pip install -U huggingface_hub
hf download yamura4/send-gate-307m

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

How to use it: You either call the gate from Python or run the included server and send messages to it. The card says the server has no login or TLS, so put a reverse proxy in front of it before exposing it to the internet.

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

Example: A chat app passes each outgoing teen message to the gate. A message with a street address and town gets block, while a message that only refuses to share an address gets allow. The card lists both kinds in its policy examples.

For newcomers and experts: If you are new to this, run the Python snippet on a few harmless messages first and read the verdicts. Remember that a gate like this is one safety layer and not a promise that everything risky is caught.

If you know the field, the decision combines four learned branches with fixed thresholds. The main classifier blocks at 0.94, the character branch at 0.95, the policy branch at 0.85, and a benign branch at 0.99 can suppress a block. The card says there are no regexes or word lists, and that dynamic inputs must not be padded.

A teen writes a message
The gate reads the text
Four learned branches score it
Thresholds decide block or allow
The app sends or stops the message

Good to know:

Measure | Result
Accuracy on 804 items | 94.2%
Harmful messages caught | 95.6%, 282 of 295
Benign messages wrongly blocked | 6.7%, 34 of 509

• Limits: the address and link branch learned from synthetic formats, so it is not a guarantee for every domain ending, spacing trick or real address style.
• Known miss: one casual phrase is a known false block, and some verdicts sit close to a threshold.
• Hardware: a 2 GB board is not recommended, and Raspberry Pi speed was not measured.
• License: MIT applies to the project code. The underlying mmBERT-base model carries its own terms, so read them before any paid use.
• Cost to run: none named.
• Data it reaches: the server downloads its five model files if they are missing and checks only the text you send it.

On Hugging Face, 14 November 2026

• Task: text classification
• Downloads in the last 30 days: 555
• Likes: 2
• License: mit
• Languages: nl, en
• Tags: onnx, student_gate_pure, text-classification, content-moderation, safety, dutch, flemish, english, mmbert, binary-classifier, no-rules, quantized

Links

• Model: https://huggingface.co/yamura4/send-gate-307m
• Files: https://huggingface.co/yamura4/send-gate-307m/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download yamura4/send-gate-307m

Then follow the usage in its model card. Nothing in this kit runs by itself.