hack-apertus-tsemin

Tse-min — Hack Apertus Track 2A (UZH): word-level semantic differences with Apertus read-back probabilities

Team Tse-min (Zhengxun Yin, University of Zurich). For an English page of a Swiss federal website and its German, French or Italian version, the system scores every word by how likely its meaning is missing or different on the other side (SwissGov-RSD). Apertus-v1.5-8B reads each document with and without the other page in the prompt; words that stay surprising are the suspects. No training; the core signal needs 4 requests per document pair.

Test split (official script): Spearman 0.323 (de .412 / fr .189 / it .367); the organisers’ CTFAlign alignment system 0.291; the paper’s prompting baseline with the same Apertus-8B −0.015.


Project root (code, data, predictions)	`track_2a/` — start with its README
Method description (2 pages)	`track_2a/Tse-min_Method.pdf`
Technical report (6 pages)	`track_2a/Tse-min_Report.pdf`
Demo (opens in the browser)	https://yinzhengxun.github.io/hack-apertus-tsemin/track_2a/demo/viewer.html
Submitted predictions	`track_2a/data/predictions/full/`
Run	`export LLM_API_KEY=...`, then `make run` here (forwards to `track_2a/`; needs Docker)

Licences: code Apache-2.0 (LICENSE); report and method description CC-BY-4.0; third-party data under their own terms (track_2a/THIRD_PARTY_NOTICES.md).

The rest of this file is the description that came with the Hack Apertus project template.

Hack Apertus — project template

Template repository for Hack Apertus submissions. Every project keeps almost the same layout, so organizers and judges find the same things in the same place.

Select your track

This repository holds one example project per track:

track_1a/
track_1b/
track_2a/
track_2b/

Keep the directory for the track you are competing in exactly as it is — don’t rename it or move its files — and delete the other track directories. That directory is your project root. Keep the files and directories as shown below.

The structure

Path	What it is
`README.md`	The challenge description and submission requirements for your track
`technical_report.md`	The deeper write-up: architecture, evaluation, limitations
`Makefile`	`make run` must spin up your project
`src/`	Your code
`data/`	Datasets — `track_1a`, `track_2a` and `track_2b` only; max. 100 MB
`findings/`	Issue files — `track_1a` only
`docs/`	Diagrams, notes, longer write-ups

Run it

Judges run make run from the root of the project, on a clean checkout:

make run

make run is expected to run the project using Docker, since that is how the judges will run it.

If you used other local open-weight models, include instructions for running the project in your technical report.

Use the following environment variables:

LLM_NAME — name and version of the model
LLM_BASE_URL — endpoint base URL
LLM_API_KEY — your API key

Getting started

Click Use this template to create your own repository.
Delete the other track directories. Don’t rename or restructure yours.
Read its README.md and fill in technical_report.md.
Make make run work from the root of the project, on a clean checkout.

License

All Hack Apertus projects are open-sourced. Please check our Terms & Conditions for specific licensing details (6. What you build is open source): https://hackapertus.ch/terms-and-conditions

This site is open source. Improve this page.