NLPnorth/under-the-influence

Under the Influence — Parsed Common Corpus (English, 10B tokens)

This repository covers a fully, automatic dependency-parsed English subset of Common Corpus, released as CoNLL-U and utilized in the work 'Under the Influence — of What? A Controlled Evaluation of Influence Functions' (Sindlinger et al., 2026). It consists of ~10B CoNLL-U tokens across 341M sentences and 2.6M documents, with the per-document licensing and provenance metadata of Common Corpus preserved on every document.

Format

We release the data as a compressed collection of CoNLL-U files, providing linguistically annotated / parsed data. For details about the file format, we refer to https://universaldependencies.org/. The dataset viewer is disabled. For details about the annotation / parsing, we refer to the paper (App. C.1).

The data is split into 46 shards of ~2 GB, each gzipped, together 90 GB compressed and 568 GB uncompressed. Shards are cut only at document boundaries, so each is independently valid CoNLL-U and no document loses its metadata header; concatenating all 46 in name order reproduces the corpus exactly.

cc-en-10b-00000-of-00046.conllu.gz
...
cc-en-10b-00045-of-00046.conllu.gz

Each document opens with its provenance block:

# newdoc id = PleIAs/common_corpus/50
# meta::identifier = US-201113336580-A_2
# meta::collection = USPTO
# meta::open_type = Open Government
# meta::license = Public Domain
# meta::date = 2011
# meta::language = English
# meta::language_type = Spoken
# sent_id = PleIAs/common_corpus/50/0
# text = The programming screen 100 also includes a mode selection control element 114 ...
1	The	the	DET	DT	Definite=Def|PronType=Art	3	det	_	_
...

The sent_id scheme is PleIAs/common_corpus/<document>/<sentence>, so every sentence traces back to its source document and thus to its licence and attribution.

Licensing

Common Corpus is a collection of public-domain and permissively licensed text, and this release preserves the original per-document licence in meta::license.

For upstream efforts on removing personally identifiable information and toxic content, see §4.4 and §4.5 of the Common Corpus paper.

Citation

If you find the work in this repository useful, please don't forget to cite:

@inproceedings{sindlinger-etal-under-the-influence,
  title     = {Under the Influence -- of What? A Controlled Evaluation of Influence Functions},
  author    = {Sindlinger, Johannes Gabriel and G{\"u}ven, Arzu Burzu and
               van der Goot, Rob and Rogers, Anna},
  year      = {2026},
}

Please also cite Common Corpus:

@article{langlais2025common,
  title   = {Common Corpus: The Largest Collection of Ethical Data for LLM Pre-Training},
  author  = {Langlais, Pierre-Carl and Rosas Hinostroza, Carlos and Nee, Mattia and
             Arnett, Catherine and Chizhov, Pavel and Jones, Eliot and Girard, Ir{\`e}ne and
             Mach, David and Stasenko, Anastasia and Yamshchikov, Ivan P.},
  journal = {CoRR},
  volume  = {abs/2506.01732},
  year    = {2025},
}