Replaces the PyTorch-only snippet with a **Use it** section:
- **Python:** `pip install millid`, then `MillID.from_pretrained(...)`, `predict` and `identify_document`, plus the `millid predict` command.
- **JavaScript:** `@baobabtech/millid`, for Node (onnxruntime-node) and browsers (onnxruntime-web).
- **PyTorch:** `millid.py`, labelled as the training code.
- **The ONNX graph:** its inputs and outputs.
All of it runs the model in-process from `onnx/model.onnx`, with no server and no PyTorch.
Everything described here is implemented and tested in https://github.com/baobab-tech/millid/pull/1:
- Python and JavaScript are tested against the same fixtures: every encoded id equal, and probabilities within 1e-4.
- Headless Chrome gave the same answers.
**Merge after that PR.** The install line points at its `main` branch, so these instructions only work once it is merged.
8 days ago
ONNX export: n-gram orders summed one at a time (a third of the peak memory)
