Join us for the DKZ.2R Data Filtering Challenge an win up to 300 EUR!

Curate the training data for a language model without seeing a single line of the text.

A language model is only as good as the data it is trained on, and deciding what to keep or discard from a raw web corpus is a large part of that work. This challenge puts that decision in your hands.

You receive a set of precomputed per-document features for a large web corpus: measures of length, formatting, language, and character composition. Working from these features alone, you design a filtering configuration that selects a “high-quality” subset of documents. We then apply your configuration, fine-tune a language model on the selected subset, and evaluate it on a held-out German benchmark. The configuration that produces the best model wins.

This challenge is open to anyone curious about how training data shapes a language model, from students to practicing NLP/ML researchers. You do not need to train a model yourself or write code beyond editing a YAML config file: the core skill is reasoning about the features (what a high ratio of short lines or a low language-detection score tells you about a document) and turning that reasoning into threshold rules. Prior exposure to text-corpus cleaning or data-quality work helps but is not required.

The baseline you have to beat replicates the filtering heuristics published in the OpenGPT-X data-processing pipeline (Brandizzi et al., 2024), a peer-reviewed data-curation method.

You can join as an individual or as a group of max. four participants and win up to 300 EUR per team!

(Prizes: first place: 300 EUR, second place 150 EUR, and third place 100 EUR)

Join the challenge on Codabench: Codabench Challenge Link

Event Details:

• Time: Friday, August 7th to Friday, September 18th.
• Location: Online on Codabench

As always, this event is free of charge!