HPLT 3.0: Deduplication Strategy Comparison Results

Dataset Description

This dataset contains fine-grained results from our HPLT 3.0 pre-release evaluations comparing different data deduplication stategies for the pre-HPLT 3.0 corpora with the previous HPLT 2.0 version. We compare the following data deduplication strategies to guide our design choices, and guard against data quality regression compared to HPLT 2.0: pre-HPLT 3.0 CD (per-crawl deduplication), pre-HPLT 3.0 GD (global deduplication), and pre-HPLT 3.0 GDR (global deduplication & rehydration). We pretrain 2.2B Llama-style decoder models on 30B tokens for each selected language and evaluate them using HPLT-E, a multilingual evaluation framework for comprehensive multi-prompt k-shot evaluation across 124 tasks and 500+ prompts in nine typologically diverse languages: Spanish (spa_Latn), French (fra_Latn), Czech (ces_Latn), Ukrainian (ukr_Cyrl), Finnish (fin_Latn), Catalan (cat_Latn), Galician (glg_Latn), Basque (eus_Latn), and Norwegian (Bokmål and Nynorsk; nor_Latn).

• Curated by: High Performance Language Technologies (HPLT)
• Languages: Spanish, French, Czech, Ukrainian, Finnish, Catalan, Galician, Basque, Norwegian Bokmål, and Norwegian Nynorsk
• Paper: arxiv.org/abs/2511.01066
• Repository: github.com/hplt-project/hplt-e
• License: Apache 2.0

Please find more details in our paper and GitHub repository.

Uses

This dataset is intended for reproducibility and research purposes. Find an example on how to access the results:

from datasets import load_dataset

dataset = load_dataset("HPLT/2505-deduplication-evals", "spa_Latn", split="results").to_pandas()

Dataset Structure

Dataset Instances

Each dataset instance looks as follows:

{
  'corpus': 'HPLT 2.0',
  'category': 'Commonsense reasoning',
  'dataset': 'xstorycloze_es',
  'task': 'xstorycloze_es_p2',
  'prompt': "{{[input_sentence_1, input_sentence_2, input_sentence_3, input_sentence_4] | join(' ') }}\n¿Qué ocurre después?\nA. {{ sentence_quiz1}} \nB. {{sentence_quiz2}}\nRespuesta:",
  'model': '1B',
  'ckpt_num': 500,
  'score': 52.813}
}

Dataset Fields

• corpus: corpus name (pre-HPLT 3.0 CD, pre-HPLT 3.0 GD, pre-HPLT 3.0 GDR, HPLT 2.0)
• category: task category
• dataset: evaluation dataset name
• task: evaluation task (refers to a specific prompt)
• prompt: prompt used for evaluation
• model: number of pretraining tokens (B)
• ckpt_num: number identifier for model
• score: standard metric performance score

Cite Us

@article{oepen2025hplt,
  title={HPLT\~{} 3.0: Very Large-Scale Multilingual Resources for LLM and MT. Mono-and Bi-lingual Data, Multilingual Evaluation, and Pre-Trained Models},
  author={Oepen, Stephan and Arefev, Nikolay and Aulamo, Mikko and Ba{\~n}{\'o}n, Marta and Buljan, Maja and Burchell, Laurie and Charpentier, Lucas and Chen, Pinzhen and Fedorova, Mariya and de Gibert, Ona and others},
  journal={arXiv preprint arXiv:2511.01066},
  year={2025}
}

Contact Us

• Vladislav Mikhailov vladism@ifi.uio.no
• Stephan Oepen oe@ifi.uio.no