Salvato in:
Dettagli Bibliografici
Autori principali: Burns, Thomas F, Parcalabescu, Letitia, Wäldchen, Stephan, Barlow, Michael, Ziegltrum, Gregor, Stampa, Volker, Harren, Bastian, Deiseroth, Björn
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:https://arxiv.org/abs/2505.00022
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866911556852252672
author Burns, Thomas F
Parcalabescu, Letitia
Wäldchen, Stephan
Barlow, Michael
Ziegltrum, Gregor
Stampa, Volker
Harren, Bastian
Deiseroth, Björn
author_facet Burns, Thomas F
Parcalabescu, Letitia
Wäldchen, Stephan
Barlow, Michael
Ziegltrum, Gregor
Stampa, Volker
Harren, Bastian
Deiseroth, Björn
contents Scaling data quantity is essential for large language models (LLMs), yet recent findings show that data quality can significantly boost performance and training efficiency. We introduce a German-language dataset curation pipeline that combines heuristic and model-based filtering techniques with synthetic data generation. We use our pipeline to create Aleph-Alpha-GermanWeb, a 628B-word German pre-training dataset composed of three subsets drawing from: (1) Common Crawl web data (organic subset; 78B words), (2) FineWeb2 (organic subset; 235B), and (3) synthetically-generated data conditioned on actual, organic web data (synthetic subset; 329B). We evaluate our dataset by pre-training both a 1B Llama-style model and an 8B tokeniser-free hierarchical autoregressive transformer (HAT) from scratch. A comparison on German-language benchmarks, including MMMLU, shows significant performance gains of Aleph-Alpha-GermanWeb over FineWeb2 alone. This advantage holds at the 8B scale even when FineWeb2 is enriched by human-curated high-quality data sources such as Wikipedia. Our findings support the growing body of evidence that model-based data curation and synthetic data generation can significantly enhance LLM pre-training datasets.
format Preprint
id arxiv_https___arxiv_org_abs_2505_00022
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Aleph-Alpha-GermanWeb: Improving German-language LLM pre-training with model-based data curation and synthetic data generation
Burns, Thomas F
Parcalabescu, Letitia
Wäldchen, Stephan
Barlow, Michael
Ziegltrum, Gregor
Stampa, Volker
Harren, Bastian
Deiseroth, Björn
Computation and Language
Artificial Intelligence
Machine Learning
Scaling data quantity is essential for large language models (LLMs), yet recent findings show that data quality can significantly boost performance and training efficiency. We introduce a German-language dataset curation pipeline that combines heuristic and model-based filtering techniques with synthetic data generation. We use our pipeline to create Aleph-Alpha-GermanWeb, a 628B-word German pre-training dataset composed of three subsets drawing from: (1) Common Crawl web data (organic subset; 78B words), (2) FineWeb2 (organic subset; 235B), and (3) synthetically-generated data conditioned on actual, organic web data (synthetic subset; 329B). We evaluate our dataset by pre-training both a 1B Llama-style model and an 8B tokeniser-free hierarchical autoregressive transformer (HAT) from scratch. A comparison on German-language benchmarks, including MMMLU, shows significant performance gains of Aleph-Alpha-GermanWeb over FineWeb2 alone. This advantage holds at the 8B scale even when FineWeb2 is enriched by human-curated high-quality data sources such as Wikipedia. Our findings support the growing body of evidence that model-based data curation and synthetic data generation can significantly enhance LLM pre-training datasets.
title Aleph-Alpha-GermanWeb: Improving German-language LLM pre-training with model-based data curation and synthetic data generation
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2505.00022