A multilingual language-modelling dataset built from Universal Dependencies 2.17 gold treebanks and UDPipe-parsed silver data. Lexemes (NOUN, VERB, ADJ, ADV, PROPN, NUM) are replaced with structured tokens encoding part-of-speech, morphological features, and a dependency-frame cluster label. Function words are kept as lowercased surface forms, namespaced by language.

Languages

ar, de, en, es, eu, fi, hi, hy, id, ja, ko, lv, ru, te, tr, zh

16 languages total: 13 with UD gold + silver data, 3 (de, ja, ru) with UD gold only.

Token format

Each sentence is one line of space-separated tokens:

Token type	Example	Meaning
Lexeme	`[en\|VERB\|frame3\|Mood=Ind\|Tense=Past]`	English verb, dependency frame 3, indicative past
Function word	`en::the`	English surface form "the"
Sentence end	`<eos>`	End of sentence marker

Language code is a two-letter ISO 639-1 code prepended to structural tokens.
frame0–frame7: dependency frame cluster (K=8 KMeans on syntactic attachment features, fitted independently per POS per language).
Morphological features are from the UD annotation, sorted alphabetically.

Splits

Split	Sentences	Notes
train	1,374,513	All training sections
dev	64,817	All dev sections
test	98,888	All test sections

Files

data/
  all_train.txt.gz        combined training set (all languages, randomised)
  all_dev.txt.gz          combined dev set
  all_test.txt.gz         combined test set
langs/
  {lang}/
    train_gold.txt.gz     UD gold training sentences for this language
    train_silver.txt.gz   UDPipe-parsed silver training sentences
    dev.txt.gz            dev sentences (UD gold only)
    test.txt.gz           test sentences (UD gold only)
dataset_info.json         per-language gold/silver/total token counts
factor_vocab.json         feature vocabulary for FactoredDelexLM
delex_vocab.txt           flat token vocabulary with counts

The langs/ files allow selecting subsets of languages or gold-only training. The data/ aggregate files are pre-shuffled across languages and sources.

Usage

Download with the HuggingFace CLI:

pip install huggingface-hub
huggingface-cli download REPO_NAME --repo-type dataset --local-dir ./delex_data
cd delex_data && gunzip data/*.gz langs/*/*.gz

Train on all languages (combined):

python train.py --data-dir delex_data/data/ --out-dir runs/run1/ \
    --factored --factor-vocab delex_data/factor_vocab.json

Train on a language subset (recombine from langs/):

cat delex_data/langs/en/train_gold.txt delex_data/langs/en/train_silver.txt \
    delex_data/langs/de/train_gold.txt > my_train.txt

Source

Universal Dependencies 2.17: https://universaldependencies.org
Silver data: UDPipe v2 parses of Common Crawl data
Dependency profiles: per-lemma syntactic attachment statistics, K=8 KMeans within POS

Downloads last month: 36

Total file size:

380 MB