README.md

Qwen3-Inspired Pre-training Dataset

Overview

This dataset is a curated mixture of high-quality text data designed for large language model pre-training, inspired by the Qwen3 methodology. The dataset includes both training and validation splits.

Dataset Statistics

Total Size: 10.42 billion tokens

• Training Split: 9.89 billion tokens (94.9%)
• Validation Split: 0.53 billion tokens (5.1%)

Data Sources (Combined)

• dclm_baseline: 5.06B tokens (48.56%) - 4,088,916 documents
• the_stack: 1.65B tokens (15.79%) - 383,490 documents
• common_corpus: 1.5B tokens (14.36%) - 381,841 documents
• mini_pile: 1.43B tokens (13.73%) - 999,858 documents
• math_pile: 0.79B tokens (7.55%) - 72,936 documents

Training Split Statistics

• dclm_baseline: 4.81B tokens (48.61%) - 3,884,088 documents
• the_stack: 1.58B tokens (15.97%) - 363,502 documents
• common_corpus: 1.42B tokens (14.37%) - 361,913 documents
• mini_pile: 1.36B tokens (13.78%) - 949,859 documents
• math_pile: 0.72B tokens (7.26%) - 68,947 documents

Validation Split Statistics

• dclm_baseline: 0.25B tokens (47.69%) - 204,828 documents
• common_corpus: 0.08B tokens (14.22%) - 19,928 documents
• math_pile: 0.07B tokens (12.89%) - 3,989 documents
• mini_pile: 0.07B tokens (12.86%) - 49,999 documents
• the_stack: 0.07B tokens (12.33%) - 19,988 documents

Data Processing Pipeline

1. Data Collection: Sourced from multiple high-quality datasets
2. Standardization: All data transformed to consistent format with text, info, and source_data fields
3. Train/Validation Split: Created 95%/5% splits within each source dataset
4. Exact Deduplication: Removed identical documents within each split
5. Near Deduplication: Used MinHashLSH with Jaccard similarity threshold of 0.85
6. Quality Filtering: Applied content-based filtering during processing
7. Shuffling: Applied shuffling within each large shard for better data distribution

Data Format

Each example contains:

• text: The main text content
• info: Metadata from the original dataset (as string)
• source_data: Source dataset identifier

Splits

The dataset contains two splits:

• train: Training data (95% of each source dataset)
• validation: Validation data (5% of each source dataset)

Tokenization

Token counts were computed using the Llama3 tokenizer (meta-llama/Meta-Llama-3-8B).

Usage

from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("bluelightai-dev/qwen_clt_pretrain_data")

# Load specific splits
train_dataset = load_dataset("bluelightai-dev/qwen_clt_pretrain_data", split="train")
val_dataset = load_dataset("bluelightai-dev/qwen_clt_pretrain_data", split="validation")

Dataset Sources

The dataset combines data from the following sources:

• DCLM Baseline: High-quality web text from DataComp-LM
• Common Corpus: Multilingual web text corpus
• The Stack: Deduplicated source code
• Mini Pile: Academic and reference texts
• Math Pile: Mathematical content and reasoning datasets

License

Please refer to the individual source dataset licenses. This mixture is provided for research purposes.

Citation

If you use this dataset, please cite the original source datasets and this work.