license: cc-by-4.0
language:

• ara
• ben
• bos
• bul
• cat
• ces
• dan
• deu
• ell
• eng
• est
• eus
• fas
• fin
• fra
• gle
• glg
• gsw
• heb
• hin
• hrv
• hun
• ind
• isl
• ita
• jpn
• kat
• kor
• lat
• lav
• lit
• ltg
• mkd
• mlt
• nld
• nno
• nob
• pol
• por
• ron
• rus
• slk
• slv
• spa
• sqi
• srp
• swa
• swe
• tha
• tur
• ukr
• urd
• vie
• yue
• zho
  tags:
• propella
• data
• annotation
• filtering
• curation
• quality
• fineweb
• finepdfs
• nemotron
• german-commons
• metadata
  pretty_name: propella annotations
  size_categories:
• 1B<n<10B
  configs:
• config_name: fineweb-2
  default: true
  data_files:
  • split: all
    path:
    • data/propella-1-4b/fineweb-2//.parquet
  • split: deu_Latn
    path:
    • data/propella-1-4b/fineweb-2/deu_Latn/*.parquet
  • split: fin_Latn
    path:
    • data/propella-1-4b/fineweb-2/fin_Latn/*.parquet
  • split: fra_Latn
    path:
    • data/propella-1-4b/fineweb-2/fra_Latn/*.parquet
  • split: ita_Latn
    path:
    • data/propella-1-4b/fineweb-2/ita_Latn/*.parquet
  • split: spa_Latn
    path:
    • data/propella-1-4b/fineweb-2/spa_Latn/*.parquet
  • split: swe_Latn
    path:
    • data/propella-1-4b/fineweb-2/swe_Latn/*.parquet
• config_name: finepdfs
  data_files:
  • split: all
    path:
    • data/propella-1-4b/finepdfs//.parquet
  • split: ces_Latn
    path:
    • data/propella-1-4b/finepdfs/ces_Latn/*.parquet
  • split: dan_Latn
    path:
    • data/propella-1-4b/finepdfs/dan_Latn/*.parquet
  • split: deu_Latn
    path:
    • data/propella-1-4b/finepdfs/deu_Latn/*.parquet
  • split: eng_Latn
    path:
    • data/propella-1-4b/finepdfs/eng_Latn/*.parquet
  • split: fin_Latn
    path:
    • data/propella-1-4b/finepdfs/fin_Latn/*.parquet
  • split: fra_Latn
    path:
    • data/propella-1-4b/finepdfs/fra_Latn/*.parquet
  • split: ita_Latn
    path:
    • data/propella-1-4b/finepdfs/ita_Latn/*.parquet
  • split: hun_Latn
    path:
    • data/propella-1-4b/finepdfs/hun_Latn/*.parquet
  • split: nld_Latn
    path:
    • data/propella-1-4b/finepdfs/nld_Latn/*.parquet
  • split: nob_Latn
    path:
    • data/propella-1-4b/finepdfs/nob_Latn/*.parquet
  • split: pol_Latn
    path:
    • data/propella-1-4b/finepdfs/pol_Latn/*.parquet
  • split: por_Latn
    path:
    • data/propella-1-4b/finepdfs/por_Latn/*.parquet
  • split: ron_Latn
    path:
    • data/propella-1-4b/finepdfs/ron_Latn/*.parquet
  • split: spa_Latn
    path:
    • data/propella-1-4b/finepdfs/spa_Latn/*.parquet
  • split: swe_Latn
    path:
    • data/propella-1-4b/finepdfs/swe_Latn/*.parquet
• config_name: hplt-3
  data_files:
  • split: all
    path:
    • data/propella-1-4b/hplt-3//.parquet
  • split: fin_Latn
    path:
    • data/propella-1-4b/hplt-3/fin_Latn/*.parquet
  • split: deu_Latn
    path:
    • data/propella-1-4b/hplt-3/deu_Latn/*.parquet
• config_name: finewiki
  data_files:
  • split: all
    path:
    • data/propella-1-4b/finewiki/*.parquet
• config_name: SYNTH
  data_files:
  • split: all
    path:
    • data/propella-1-4b/SYNTH/*.parquet
• config_name: nemotron-cc
  data_files:
  • split: all
    path:
    • data/propella-1-4b/nemotron-cc//.parquet
  • split: high_actual
    path:
    • data/propella-1-4b/nemotron-cc/high-actual/*.parquet
• config_name: nemotron-cc-10k-sample
  data_files:
  • split: all
    path:
    • data/propella-1-4b/nemotron-cc-10k-sample/*.parquet
• config_name: german-commons
  data_files:
  • split: all
    path:
    • data/propella-1-4b/german-commons/*.parquet

This dataset contains document annotations produced with propella-1-4b, a small multilingual LLM that annotates text documents across six categories: core content, classification, quality & value, audience & purpose, safety & compliance, and geographic relevance. The annotations can be used to filter, select, and curate LLM training data at scale.

Properties

Each document is annotated across 18 properties organized into six categories:

Read the property reference for detailed definitions and enum values.

Dataset Overview

This dataset is work-in-progress.
We plan to add lots of annotations over time.
Want to suggest a dataset to be annotated next?
Feel free to open a discussion in the community tab!
Want to contribute significant compute for more annotations?
Get in touch.

Currently, we provide annotations for the following datasets:

fineweb-2
Source: FineWeb-2

finepdfs
Source: FinePDFs

hplt-3
Source: HPLT3.0

finewiki
Source: finewiki

SYNTH
Source: PleIAs/SYNTH
Note: text = f"{row['query']}\n\n{row['synthetic_reasoning']}\n\n{row['synthetic_answer']}

nemotron-cc
Source: Nemotron-CC
Note: This is only a subset of the high-actual-actual split.

nemotron-cc-10k-sample
Source: nemotron-cc-10K-sample
A sample from nemotron-cc, containing 10k documents from each of the five quality categories.

german-commons
Source: German Commons

Usage

import datasets as hfds

load annotations for German FineWeb-2
annotations = hfds.load_dataset("openeurollm/propella-annotations", "fineweb-2", split="deu_Latn")
example filter: high educational value
high_edu_ids = set(
    annotations
    .filter(lambda x: x["educational_value"] == "high")
    ["id"]
)
filter German FineWeb-2 by matching ids
ds = hfds.load_dataset("HuggingFaceFW/fineweb-2", "deu_Latn", split="train", streaming=True)
filtered = ds.filter(lambda x: x["id"] in high_edu_ids)
for doc in filtered:
    print(doc["text"][:500])
    break

License

The annotation data in this repository is released under the CC-BY-4.0 license.

Citation

TBA

Acknowledgements

• This project used compute resources made available via the EuroHPC Joint Undertaking (EuroHPC JU) AI Factories initiative (AI for Industrial Innovation – Large Scale Access) on the EuroHPC supercomputer LEONARDO operated by CINECA and the LEONARDO consortium.
• This project used compute resources made available via the EuroHPC Joint Undertaking (EuroHPC JU) AI Factories initiative (AI for Industrial Innovation – Large Scale Access) on the EuroHPC supercomputer MareNostrum 5 operated by the Barcelona Supercomputing Center (BSC).
• This project is supported by the OpenEuroLLM project, co-funded by the Digital Europe Programme under GA no. 101195233. For more information see openeurollm.eu.
• This project is supported by the LLMs4EU project, co-funded by the Digital Europe Programme under GA no. 101198470. For more information see LLMs4EU website.
• ellamind is supported by the German Federal Ministry for Economic Affairs and Energy (BMWE) under the soofi (Sovereign Open Source Foundation Models for European Intelligence) project.
• ellamind thanks the AI Service Center for Sensitive and Critical Infrastructures (KISSKI), operated by GWDG, for additional compute access.