README.md

metadata

language:
  - bg
license: apache-2.0
task_categories:
  - text-generation
  - question-answering
  - translation
pretty_name: Bulgarian Corpus 33B
size_categories:
  - 10B<n<100B
tags:
  - bulgarian
  - llm
  - foundation-model
  - pretraining
  - sft
  - fineweb
  - science
configs:
  - config_name: pretrain
    data_files: pretrain/*.parquet
  - config_name: sft
    data_files: sft/*.parquet

Lumees Bulgarian Corpus (BG-Corpus-33B)

Dataset Summary

The Bulgarian Corpus 33B is a massive-scale, deduplicated, and cleaned dataset designed for training Foundation Models in Bulgarian. Comprising approximately 33.4 Billion tokens (measured with Qwen 2.5/Llama-3 tokenizer), it represents one of the largest open-source resources for Bulgarian LLM pretraining.

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

1. Pretrain Subset (~29.3B Tokens): A diverse mix of high-quality web data, encyclopedic knowledge, and scientific abstracts.
2. SFT Subset (~4.1B Tokens): A curated collection of instruction-following, chat, and multitask data, strictly filtered to remove alignment artifacts.

Training Recommendation: With ~33B unique high-quality tokens, we recommend training for 3 Epochs over the pretrain subset to achieve optimal convergence for models in the 7B-8B parameter range (effectively ~90B training tokens).

---

Dataset Statistics

Estimates based on Qwen 2.5 / Llama-3 Tokenization.

Subset	Format	File Type	Documents	Token Count
Pretrain	Universal Schema	Parquet (Snappy)	26,278,393	~29.31 Billion
SFT	ChatML	Parquet (Snappy)	8,663,195	~4.11 Billion
Total	-	-	34,941,588	~33.42 Billion

---

Data Structure

1. Pretraining Subset (pretrain)

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

Column	Type	Description
id	string	Unique identifier (vital for tracking).
text	string	The cleaned, deduplicated content.
source	string	Origin dataset (e.g., fineweb-2, bpos_science).
language	string	ISO Code (bg).
meta	string	Original metadata (URL, date, title, DOI) serialized as a JSON string.

2. SFT Subset (sft)

Optimized for "Instruction Pretraining" or Fine-Tuning (Axolotl/LLaMA-Factory compatible).

Column	Type	Description
messages	list	Standard OpenAI/ChatML format: [{"role": "user", ...}, {"role": "assistant", ...}]
source	string	Origin task (e.g., aya_collection, xp3x).

---

Data Composition

This corpus was built using a Quality-First strategy, blending massive web scale with high-density scientific and encyclopedic data.

Source	Type	Usage Phase	Description
FineWeb-2 (Bulgarian)	Web Crawl	Pretrain	The backbone of the corpus (cleaned web text).
FineWiki BG	Knowledge	Pretrain	Full Bulgarian Wikipedia dump with rich metadata.
BPOS (Open Science)	Scientific	Pretrain	4,700+ Titles and Abstracts from the Bulgarian Portal for Open Science (High density domain knowledge).
Aya Collection	Instruction	SFT	High-quality multilingual instruction following.
xP3x	NLP Tasks	SFT	Massive multitask dataset (Filtered for quality).
Alpaca Dictionary BG	Linguistic	SFT	Definitions, synonyms, and linguistic tasks.

---

Processing Pipeline

This dataset was engineered for Foundation Model training standards:

1. Normalization: Multiple raw data sources were mapped to a single unified schema.
2. PII Sanitization:
   • Regex Cleaning: Automated removal of Email addresses, IPv4 addresses, and Bulgarian phone numbers (e.g., +359..., 088...).
3. DB-Assisted Deduplication:
   • Exact deduplication (MD5 hashing) was performed across the entire collection.
   • Priority Strategy: High-quality sources (Wiki/Science) were processed first to claim ownership of duplicate text, ensuring the highest quality version is kept.
4. Quality Filtering (SFT):
   • The SFT subset was scrubbed of "poison" rows (e.g., where the assistant replies "None", "null", or refuses to answer due to alignment errors).
5. Sharding: Data is split into ~200k row Parquet shards for optimal download and streaming speeds.

Limitations

• Web Bias: A significant portion of the data (FineWeb) comes from the open internet and may reflect societal biases found in Bulgarian web content.
• Translation Artifacts: Some SFT data is machine-translated or aligned; while we filtered obvious errors, some translation artifacts may remain.

---

Citation & Attribution

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

@misc{bulgariancorpus33b,
  author = {Hasan KURŞUN, Kerem Berkay YANIK},
  publisher = {Lumees AI},
  title = {Bulgarian Corpus 33B},
  year = {2025},
  publisher = {HuggingFace Community},
  howpublished = {\url{[https://lumees.io](https://lumees.io)}},
  email = {[email protected]}
}