provost-model

The tiny language model behind provost's --model, packaged on its own so provost itself stays small.

provost turns command output into tidy datasets. When its parsers can't read a capture cleanly, --model asks a learned layout model which lines are headers, rows, key/value pairs or sections, and where each field starts. Every value is still cut from the input; the model only decides where. provost ships small decision trees for this. This package carries a better reader: a pretrained language model fine-tuned on provost's own labels.

Install

pip install provost[lm]          # provost, numpy and this package

or, beside an existing provost:

pip install provost-model        # pulls in numpy

provost finds it on its own: with provost-model installed and PROVOST_LAYOUT_LM unset, provost --model labels with this model instead of the trees.

• PROVOST_LAYOUT_LM=0 keeps the trees.
• PROVOST_LAYOUT_LM=DIR uses another model directory.
• provost runs the model under the Python it was installed with; when that differs, set PROVOST_PYTHON to the interpreter that has this package.

No torch, transformers or GPU is needed: provost runs the model with numpy alone (extract/model/lm_numpy.py).

What's in it

provost_model.MODEL_DIR is a directory holding:

The model is SmolLM2-135M (Apache-2.0):

• fine-tuned with a LoRA adapter and two small heads on Claude's labels of provost's 3,674-capture corpus;
• cut to its first 18 of 30 layers, with the adapter merged in;
• stored at 8 bits per weight, one scale per row.

provost's ml/export_lm.py builds this directory from a model ml/train_lm.py trains.

How good it is

Per line, five-fold cross-validation on held-out tool families. The score is how often a line's field split matches the reference label (14,468 lines in all):

The 8-bit export matches the float model:

• on fold 0, the 30-layer model scores 0.863 in float and 0.863 exported;
• on provost's round-4 benchmark (107 captures, rated blind), this model rates exactly as its float original does.

What it costs to run: about 280 MB of memory and a 0.5 s load. It takes under a second per capture on 4 CPU cores, and longer for a capture of hundreds of long lines.

On whole captures it is a second reader, not a replacement for provost's parsers. That is why provost only asks it about a capture the quality gate would otherwise hold.

Details: provost's docs/lm-2026-10-29-tiny-labeller.md.

Releasing

.github/workflows/release.yml publishes to PyPI by Trusted Publishing, with no API token, when a GitHub release is published or the workflow is run by hand.

• Set it up once on PyPI: project provost-model, owner MartinGallagher-code, repository provost_model, workflow release.yml, environment pypi.
• For each release:
  1. bump __version__ in src/provost_model/__init__.py and add a CHANGELOG.md entry;
  2. merge;
  3. publish a GitHub release tagged vX.Y.Z.

Licence

Apache-2.0, as the base model is; see LICENSE and NOTICE. provost itself is GPL-3.0-or-later.