TitleBERT for the M5 Pro Mac

A small program that reads a short string and decides whether it is a job title, something else (a name, a company, a department, a phone, junk, …), or not sure.

This folder is self-contained. Clone it, install once, train once, then score your own files. Training data is already inside.

Repo (public): https://git.binary.ovh/brahm/titlebert-mac

Production machine this is built for: Apple M5 Pro, 48 GB unified RAM. On that Mac it uses the 2025 multilingual encoder mmBERT-base (307 million knobs, 1,800+ languages).

If you only remember three commands after install:

source .venv/bin/activate
make all-mac          # check, test, prepare, train, eval, verify
make score FILE=incoming/your.csv

Start here

1. Read docs/MAC-GUIDE.md — copy, install, train, read the scores, then score production data in a dev folder.
2. Read docs/GOALS.md — what precision, recall, and validity mean, and the pass/fail floors.
3. Skim docs/IMPROVEMENTS.md — what was wrong in the lab project and what this repo fixes.
4. Skim docs/DESIGN.md — why 307M on this Mac, and why more RAM still will not do a billion rows in 12 hours.

What “done” looks like

Rung	Command that proves it	You may then…
Code works	make test + make verify-setup	Train
Lab ready	make verify — all required gates PASS	Score a labeled slice of your file
Dev ready	Your labeled slice hits the same floors	Score the full extract
Production ready	abstain goes to a person; you spot-check accepts	Load accepts into the system

This repo automates the first two rungs. The last two need your file and a person.

Clone (public)

git clone https://git.binary.ovh/brahm/titlebert-mac.git
cd titlebert-mac
make install
source .venv/bin/activate
make check

make check must print profile: mac-m5-pro and metal_mps True.

What this is not

• Not a chatbot.
• Not a dictionary of official occupations (that older tool maxed out at ~33% of real titles).
• Not proven on your languages until you score a labeled slice of your file. The encoder is multilingual; the hard exam in this folder is still mostly English messy titles plus official European titles.
• Not a billion-row-in-12-hours engine. On the M5 Pro, think millions per night, not billions.

License

Use inside your company as you like. The encoder weights you download (jhu-clsp/mmBERT-base) are MIT. Official occupation lists (ESCO, O*NET) keep their original public licences.