mimedefang/gliclass repository overview

⁠mimedefang/gliclass

mimedefang-gliclass-server keeps a GLiClass⁠ zero-shot classifier loaded in memory and answers classification requests from MIMEDefang⁠'s Mail::MIMEDefang::ML module (gliclass backend). It classifies mail from the label texts alone, in a single forward pass. You can also serve a model fine-tuned on your own mail with MIMEDefang's gliclass-train.

⁠Tags

• latest, vX.Y: built from a MIMEDefang release tag, once one exists

⁠Run

docker run -d --name gliclass -p 127.0.0.1:8688:8688 \
    -v mdml-hf:/var/lib/MIMEDefang/ml/huggingface \
    mimedefang/gliclass:latest

On the first start the model is downloaded from Hugging Face into the volume. Check the server with curl http://127.0.0.1:8688/health.

To serve your own fine-tuned model, mount it and point GLICLASS_MODEL at it:

docker run -d --name gliclass -p 127.0.0.1:8688:8688 \
    -v /srv/my-gliclass-model:/model:ro -e GLICLASS_MODEL=/model \
    mimedefang/gliclass:latest

⁠Configure MIMEDefang

use Mail::MIMEDefang::ML;
$Mail::MIMEDefang::ML::Config{backend}              = 'gliclass';
$Mail::MIMEDefang::ML::Config{gliclass}{server_url} = 'http://127.0.0.1:8688';

⁠Environment

⁠Security

The server has no authentication, and every request carries message content. Publish the port on 127.0.0.1, or firewall it so that only your MIMEDefang hosts can reach it. Never expose it to the Internet.

Source: https://github.com/The-McGrail-Foundation/MIMEDefang-Docker⁠