metadata

language:
  - de
  - es
  - fr
  - pt
  - it
  - nl
  - el
  - pl
  - cs
  - sk
task_categories:
  - text-generation
pretty_name: Occiglot Fineweb v1.0
size_categories:
  - 10B<n<100B
extra_gated_prompt: >-
  By filling the form below I understand that occiglot-fineweb is a derivative
  collection of multiple datasets which use individual licenses and their
  respective terms and conditions apply.I understand that all uses of the
  textual content in occiglot-fineweb are subject to the terms os use. I
  understand that reusing the textual content in occiglot-fineweb might not be
  legal in all countries/regions and for all use cases. I understand that 
  occiglot-fineweb is mainly targeted towards researchers and meant to be used
  in research. Occiglot reserves the right to revoke my access to this data.
  Occiglot reserves the right to modify this data at any time in accordance to
  take down requests.
extra_gated_fields:
  Name: text
  Email: text
  Affiliation: text
  Country: text
  Usecase: text
  I have explicitly checked that downloading occiglot-fineweb is legal in my jurisdiction, in the country/region where I am located right now, and for the use case that I have described above, I have also read and accepted the relevant Terms of Use: checkbox

Occiglot Fineweb v1.0

We present a more mature version of the multilingual Occiglot Fineweb corpus. In this early form, the dataset contains roughly 430M heavily cleaned documents from 10 languages. Occiglot Fineweb builds on our existing collection of curated datasets and pre-filtered web data. Subsequently, all documents were filtered with language-specific derivatives of the fine-web processing pipeline and different levels of depuplicated.

We provide the data at 3 levels of processing:

After filtering
After local deduplication (within data sources)
After global deduplocation (for each language)

We are actively working on extending this dataset with more data and further languages. For more information please refer to our blog post or join our Discord server.

Unfortunately, some of the datasets we used do not allow for re-distribution. Consequently, we had to exclude those from this version of our dataset. We are exploring different avenues to make this data available to the public as well.

Datasources

We mainly relied on two sources of data.

1. LLM-Dataset

From LLM-Datasets we took all available datasets for the considered languages (excluding OSCAR). This collection of data for LLM training is curated from various sources and contains multiple high-quality datasets.

2. Web-Data

We sourced web-crawled data from our Community-Oscar dataset.

Filtering

All data was rigorously filtered using language-specific pipelines built upon Huggingface's fine-web filters. In addition to some minor hyper-parameter adjustments we mainly modified 3 aspects to ensure language-specific quality filtering.

Adjust average-word length filters according to lingusitic characteristics of each language
Add language-specific stop words
Add a language-specific policy filter for policy and cookie filtering

Compared to the our prior version, we improved the configuration of the filtering settings, cleaned up the encoding of every document using ftfy and ran an additional language id filtering step for datasources from countries with multiple official languages (e.g. Belgium).

Deduplication

We performed minhash deduplication on all data of each language.

Importantly, we always retain the duplicate not contained in the web-crawled data for the globally deduplicated dataset. For example, if a wikipedia page is also contained in OSCAR, we drop the OSCAR duplicate, thus keeping the wikipedia subset complete. This dataset structure allows to reliably over- or undersample the custom subsets.

Statistics

For the global deduplciated set:

Language	lang-code	# Documents	# Tokens (Llama-3)
German	de	82.60M	135.46B
Spanish	es	91.89M	108.15B
French	fr	61.80M	87.61B
Portugese	pt	46.97M	54.87B
Italian	it	37.14M	58.24B
Dutch	nl	29.00M	33.78B
Greek	el	17.55M	24.21B
Polish	pl	21.43M	35.35B
Czech	cs	38.98M	25.23B
Slovak	sk	4.18M	11.13B

Total		431.53M	574.03B

Acknowledgements

The dataset creation by a compute grant at the 42 supercomputer which is a central component in the development of hessian AI, the AI Innovation Lab (funded by the Hessian Ministry of Higher Education, Research and the Art (HMWK) & the Hessian Ministry of the Interior, for Security and Homeland Security (HMinD)) and the AI Service Centers (funded by the German Federal Ministry for Economic Affairs and Climate Action (BMWK)). Some preliminary computations were conducted on the DFKI Pegasus Cluster. Parts of the preliminary data curation were funded by the German Federal Ministry for Economic Affairs and Climate Action (BMWK) through the project OpenGPT-X (project no. 68GX21007D).