AstrLink Guard

AstrLink Guard

Finds personal information and credentials in Chinese, English and mixed text, code, logs and AI coding assistant requests, locally in ONNX. For developers who mask sensitive data before it leaves a machine.

Made by QuantumNous 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: AstrLink Guard is a compact model that picks out private details and secrets from text in Chinese, English or both. It also reads source code, configuration files, logs and the requests that AI coding assistants send to language models. It runs locally in ONNX Runtime on a CPU, and the FP16 build suits a GPU. For each hit it gives the span, a label and a score, and your own code decides what to hide.

Who it suits: Developers who want to hide secrets and private details before text reaches a language model or a log, and who prefer a local check.

Industry | Who uses it | How they use it | Example
Software | Tool builders | Scan coding assistant requests for secrets | Find an AWS access key inside a prompt
Privacy | Data protection engineers | Mask personal details in logs and documents | Redact names, accounts and ID numbers
Enterprise IT | Platform teams | Add a local filter in front of a model gateway | Block credentials from leaving the company

Anyone who needs a detector for languages other than Chinese and English can skip it.

Where to use it:

• ONNX Runtime on a CPU, with the INT8 folder.
• A GPU runtime, with the FP16 folder.
• AstrLink, the app that uses this model for privacy detection.
• Windows, macOS or Linux: the download steps run on all three with Python.

Use case | Who | What they get back
Secret scanning | Developers | Spans labeled as credentials, with scores
Privacy masking | Privacy engineers | Spans for personal data, ready to redact
Request filtering | Platform teams | A per-request list of sensitive spans

How to set it up: What you need first:

• Python with pip.
• ONNX Runtime, for the CPU or your GPU.
• A tokenizer that matches the model, an XLM-R SentencePiece one.
• Your choice of folder: int8, fp32 or fp16.

1. Install the Hugging Face hub package with pip.
2. Download the repository with the hf command.
3. Pick a format folder and load its ONNX file in ONNX Runtime.
4. Feed it tokenized text and read back the labeled spans.

pip install -U huggingface_hub
hf download QuantumNous/astrlink-guard

How to use it: Send text in windows of up to 512 tokens, with a 128-token overlap for longer inputs. The model returns spans with labels and scores, and you apply the recommended minimum span score from the full card before you mask anything. To keep the older date behavior, pin the repository to revision v0.1.1.

Title: Formats in the repository
Tool | What it does
int8 folder | Dynamic INT8 build, recommended on CPU, 36 MiB
fp32 folder | FP32 reference build with higher accuracy, 141 MiB
fp16 folder | FP16 build for GPU runtimes, which takes exactly 512 tokens per input

Example: You pass in an AI coding assistant request that contains an AWS access key and a person's birth date. The model returns a span for the key and one for the date, each with a label and a score. Your code then masks both.

For newcomers and experts: Think of it as a highlighter for private details. It marks the parts of a text that look like secrets or personal data, and you choose what to hide. A good first step is to try the INT8 build on a short sample.

The model is a BertForTokenClassification with 12 layers and hidden size 384, about 36.9M parameters, based on Multilingual-MiniLM. Version 0.1.2 changed the date label to cover only dates about a specific person, which cut false alarms but lowered recall on personal dates.

Split long text into 512-token windows
Run the ONNX model on each window
Collect spans with labels and scores
Mask the spans your app cares about

Good to know: In version 0.1.2 the date label stopped firing on non-personal dates. This chart uses 7,743 such dates in coding assistant requests.

Title: Non-personal dates flagged, INT8, out of 7,743
Version 0.1.1: 7435 dates
Version 0.1.2: 21 dates

• Accuracy: general text F1 is 97.46 with INT8, 97.35 with FP32, and FP16 lands within 0.03 F1 of FP32.
• Trade-off: version 0.1.2 finds 81.5 percent of personal dates with INT8, down from 98.6 percent.
• Labels: 10 labels in all, including private_date, common_secret and account.
• License: apache-2.0
• Cost to run: no API key, account or paid service named.
• Data it reaches: files you download from Hugging Face, and the text you pass in, which stays on your machine.
• Not for: languages other than Chinese and English, or setups that must mask every date.

On Hugging Face, 13 November 2026

• Task: token classification
• Library: onnx
• Downloads in the last 30 days: 408
• Likes: 29
• License: apache-2.0
• Languages: zh, en
• Based on: microsoft/Multilingual-MiniLM-L12-H384
• Tags: onnx, bert, privacy, secret-detection, token-classification

Links

• Model: https://huggingface.co/QuantumNous/astrlink-guard
• Files: https://huggingface.co/QuantumNous/astrlink-guard/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download QuantumNous/astrlink-guard

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