lumees/dutch-corpus-200b

Lumees Dutch Corpus 200B (LDC-200B)

Dataset Summary

The Lumees Dutch Corpus 200B (LDC-200B) is the largest open-source, deduplicated, and professionally cleaned dataset designed for training Foundation Models in the Dutch language. Comprising approximately 202 Billion tokens (measured with Qwen 2.5 tokenizer), it bridges the gap between high-resource English models and the Dutch ecosystem.

The dataset is engineered for a two-stage training pipeline:

1. Pretrain Subset (~195B Tokens): A massive, diverse mix of native Dutch web data, government publications, court verdicts, and high-fidelity translated books for continual pretraining.
2. SFT Subset (~7.25B Tokens): A high-density collection of instruction-following, multi-turn chat, and reasoning tasks, including the complete Aya and xP3x Dutch collections.

---

Dataset Statistics

Estimates based on Qwen 2.5 (Byte-Level BPE) Tokenization.

Subset	Format	File Type	Token Count
Pretrain	Universal Schema	Parquet (ZSTD)	~194.95 Billion

Data Structure

1. Pretraining Subset (pretrain)

Optimized for high-throughput streaming with libraries like datatrove or nanotron.

Column	Type	Description
id	string	Unique UUIDv4 (Vital for deduplication tracking).
text	string	The cleaned, deduplicated content.
source	string	Origin dataset (e.g., fineweb-2, rechtspraak_verdicts).
language	string	ISO Code (nl).
meta	string	Original metadata (URL, date, ECLI, title) serialized as JSON string.

2. SFT Subset (sft)

Optimized for "Instruction Pretraining" or large-scale Fine-Tuning.

Column	Type	Description
id	string	Unique UUIDv4.
messages	list	Standard Conversational format: [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]
source	string	Origin task (e.g., no_robots_dutch, aya_collection).
meta	string	Task type, original ID, or template name serialized as JSON.

---

Data Composition

This corpus uses a High-Quality Native + Synthetic Transfer strategy. We prioritize native formal Dutch (Legal/Gov) and supplement it with high-quality translations for reasoning.

Pretrain Sources

Source	Type	Description
FineWeb-2 (Dutch)	Web Crawl	High-quality web backbone (subset nld_Latn).
Rechtspraak	Legal	Dutch court verdicts. Gold standard for formal logic and grammar.
FinePDFs (Dutch)	Documents	Technical papers, reports, and long-form PDF content.
FineWiki NL	Knowledge	Full Dutch Wikipedia dump (cleaned).
Officiële Publicaties	Government	Official gazettes and parliamentary papers.
Apollo Books (UMCU)	Literature	Public domain books translated via Gemini 1.5 Flash (Long Context).

SFT Sources

Source	Type	Description
BramVanroy Collection	Chat/Instruct	Includes no_robots, ultrachat_200k, dolly-15k, and alpaca (Dutch).
Aya Collection	Multilingual	Massive human-annotated and templated instruction dataset (Dutch subset).
xP3x	NLP Tasks	Cross-lingual generalization tasks (nld_Latn).

---

Processing Pipeline

This dataset was engineered using a memory-efficient Polars Streaming pipeline designed for low-resource environments:

1. Normalization: 8+ distinct data sources were mapped to the unified lumees schema.
2. Hashed Binning Deduplication:
   • To handle >100GB of data without OOM errors, data was split into 50 hashed bins.
   • Global deduplication was performed on the exact text content (whitespace normalized) within each bin.
3. Dutch PII Redaction:
   • BSN (Burgerservicenummer): Conservative regex for 9-digit identifiers.
   • IBAN: Dutch bank account numbers (NLxx...).
   • Phone: Dutch mobile and landline formats (+31, 06).
   • Postcode: Dutch format (1234 AB).
4. Sharding: Data is split into optimized Parquet shards for distributed training.

Limitations

• Translation Bias: A portion of the "reasoning" data (Books, Reasoning) is synthetic/translated. While high quality (Gemini/Tower), it may contain translation artifacts.
• Web Bias: FineWeb-2 data reflects the general internet and may contain societal biases inherent to web text.

---

Citation & Attribution

If you use this dataset in your research or product, please cite:

@misc{lumees2025dutch200b,
  author = {Hasan KURŞUN, Kerem Berkay YANIK},
  title = {Lumees Dutch Corpus 200B},
  year = {2025},
  publisher = {Lumees AI},
  howpublished = {\url{https://lumees.io}},
  email = {[email protected]}
}