propella-1: Multi-Property Document Annotation for LLM Data Curation at Scale
Paper • 2602.12414 • Published • 4
Datasets:
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Tell us the research institution and the intended use so we can follow how these annotations are used.
Log in or Sign Up to review the conditions and access this dataset content.
Propella annotations
Document id plus the 18 propella fields. Source text is not included. Join on id.
Annotated with ellamind/propella-1-4b @ bf607e62b6afa3e0e8d71c4d08d1429d9a09c82f. Schema follows openeurollm/propella-annotations. Documents were truncated at 50,000 characters with a trailing <truncated_content> line.
CLASSLA_web_2
| split | annotations | source |
|---|---|---|
| sl | 4,790,039 | zID4si/CLASSLA_web_sl_2_enriched default @ 73424c3ffeb4456d69fb812126532e9989bf2de2 |
Source documents stay in CLASSLA-web 2.0, CLARIN.SI http://hdl.handle.net/11356/2079. The filtered CPT subset of that corpus uses the same ids and is not annotated separately. List fields: content_type, business_sector, technical_content, regional_relevance, country_relevance.
Key facts
Full coverage: 4,790,039 / 4,790,039.
| metric | share |
|---|---|
| Complete or mostly complete | 89.22% |
| Complete or mostly content | 98.84% |
| Excellent or good quality | 72.16% |
| High or moderate educational value | 13.02% |
| Safe | 95.82% |
| No PII | 93.58% |
| Non-technical label present | 72.95% |
| European label present | 83.72% |
| Mentions Slovenia | 74.41% |
Full distributions, crosstabs, and charts: classification report and PDF. Chart images are in CLASSLA_web_2/report/figures.
fineweb-2
| split | annotations | source |
|---|---|---|
| sl | 12,091,577 | HuggingFaceFW/fineweb-2 slv_Latn @ af9c13333eb981300149d5ca60a8e9d659b276b9 |
Source documents stay in FineWeb-2 slv_Latn @ af9c13333eb981300149d5ca60a8e9d659b276b9 under ODC-By. Train rows come before test rows. No language-score filter is applied. The source corpus is ODC-By; these rows are annotations only. List fields: content_type, business_sector, technical_content, regional_relevance, country_relevance.
Key facts
Full coverage: 12,091,577 / 12,091,577.
| metric | share |
|---|---|
| Complete or mostly complete | 89.43% |
| Complete or mostly content | 91.92% |
| Excellent or good quality | 73.61% |
| High or moderate educational value | 12.30% |
| Safe | 97.58% |
| No PII | 93.67% |
| Non-technical label present | 76.26% |
| European label present | 82.86% |
| Mentions Slovenia | 73.63% |
Full distributions, crosstabs, and charts: classification report and PDF. Chart images are in fineweb-2/report/figures.
Usage
import datasets as hfds
annotations_classla = hfds.load_dataset("zID4si/propella-annotations", "CLASSLA_web_2", split="sl")
annotations_fineweb_2 = hfds.load_dataset("zID4si/propella-annotations", "fineweb-2", split="sl")
License
Annotation rows are CC-BY-4.0. Source documents stay in their own repositories.
Citation
If you use propella-annotations in your research, please cite the dataset, the source corpus, and the annotation model.
BibTeX
@misc{propellaannotations2026,
title = {propella-annotations},
author = {Zavod za informacijsko družbo (zID)},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/zID4si/propella-annotations}}
}
@misc{kuzman2026classlaweb,
title = {South Slavic web corpus collection CLASSLA-web 2.0},
author = {Kuzman Pungeršek, Taja and Rupnik, Peter and Ljubešić, Nikola},
year = {2026},
howpublished = {Slovenian language resource repository CLARIN.SI},
note = {http://hdl.handle.net/11356/2079}
}
@misc{idahl2026propella1,
title = {propella-1: Multi-Property Document Annotation for LLM Data Curation at Scale},
author = {Maximilian Idahl and Benedikt Droste and Björn Plüster and Jan Philipp Harries},
year = {2026},
eprint = {2602.12414},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2602.12414}
}
APA
Zavod za informacijsko družbo (zID). (2026). propella-annotations [Dataset]. Hugging Face. https://huggingface.co/datasets/zID4si/propella-annotations
Kuzman Pungeršek, T., Rupnik, P., & Ljubešić, N. (2026). South Slavic web corpus collection CLASSLA-web 2.0. Slovenian language resource repository CLARIN.SI. http://hdl.handle.net/11356/2079
Idahl, M., Droste, B., Plüster, B., & Harries, J. P. (2026). propella-1: Multi-Property Document Annotation for LLM Data Curation at Scale. arXiv. https://arxiv.org/abs/2602.12414
FineWeb-2 slv_Latn is ODC-By: https://huggingface.co/datasets/HuggingFaceFW/fineweb-2
- Downloads last month
- 47