README.md

metadata

title: RuleChef Live
colorFrom: gray
colorTo: blue
sdk: static
app_file: index.html
license: apache-2.0
short_description: Model answers become regex rules as a stream plays
models:
  - vllm-sr/vela3-fast
datasets:
  - mattmdjaga/text-anonymization-benchmark-train

RuleChef Live

A request stream where a teacher model's answers become labelled examples and RuleChef turns them into regex rules. Rules answer the requests they cover, with no model call; the teacher (vllm-sr/vela3-fast) answers the rest.

• Task: dates and times, codes and reference numbers, and quantities in sentences from European Court of Human Rights judgments (TAB, public train split).
• Rule learning: RuleChef search synthesis (synth="search", no LLM), re-run on CPU after the first 50 labelled examples and then every 100 new ones.
• The page replays one offline run of sim.py (data/replay.json). The rules run live in the browser (engine.js, a port of RuleChef's executor and of the gate); the teacher does not run on this static page.

Files

• index.html, app.js, style.css: the page. engine.js: rule execution and gate in the browser.
• cascade.py: gate, rule execution (RuleChef's executor), scoring. sim.py: the stream run that writes data/replay.json.
• prep.py, run_teacher.py, teacher_core.py: how the requests and the teacher's answers in data/requests.jsonl were made.
• pyref.py, verify_engine.js: check that engine.js gives the same spans and gate decisions as the Python code.

Reproduce (Python 3.12; torch CPU, transformers==5.17.0, pyarrow, and RuleChef from git+https://github.com/KRLabsOrg/rulechef@8173f96f1b48cf1d66b35bc5ebe9d83be962a610), in a fresh folder work/: python prep.py <TAB parquet> work; run run_teacher.py over work/requests.jsonl into work/teacher.jsonl; python sim.py work work/replay.json.

Licences

• Code in this Space: Apache-2.0.
• RuleChef: Apache-2.0. vela3-fast: Apache-2.0.
• TAB (Text Anonymization Benchmark, Pilán et al. 2022, Norsk Regnesentral): MIT, notice in DATA_LICENSE_TAB.txt. data/requests.jsonl holds sentences from its train split (mattmdjaga/text-anonymization-benchmark-train @f0b7eeb), with TAB's labels for the three types and the teacher's answers.