propella-1: Multi-Property Document Annotation for LLM Data Curation at Scale

Maximilian Idahl 1,2 Benedikt Droste 1 Björn Plüster 1 Jan Philipp Harries 1 1 ellamind 2 Leibniz University Hannover · 2026

Read propella-1: Multi-Property Document Annotation for LLM Data Curation at Scale by Maximilian Idahl 1,2 Benedikt Droste 1 Björn Plüster 1 Jan Philipp Harries 1 1 ellamind 2 Leibniz University Hannover (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-12414.txt.

Description

Since FineWeb-Edu, most data curation for model training has relied on a single quality score from a small classifier, which hides trade-offs and blocks flexible filtering. Maximilian Idahl led this work with three co-authors, presenting propella-1, a family of small models that work in many languages, at 0.6, 1.7, and 4 billion parameters. They tag documents across 18 traits in six groups, from core content and type to safety, audience, and place relevance.

The models cover 57 languages and emit structured JSON under a fixed schema. Judged against a frontier commercial model, the 4-billion version matches it more closely than much larger general models do. The team also releases over three billion tags across major corpora, revealing sharp gaps in quality, depth of reasoning, and mix of content that single scores cannot capture, under permissive commercial-use licences.

Catalog record

CREATORMaximilian Idahl 1,2 Benedikt Droste 1 Björn Plüster 1 Jan Philipp Harries 1 1 ellamind 2 Leibniz University Hannover
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE84 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-12414

Browse the full library · About the Tower of Babel