Nj-1111/Copernicus-Dataset

Copernicus Dataset

A curated, multi-domain dataset for tokenizer training and small language-model pretraining, covering natural language, source code, and mathematics.

The dataset was created as the training corpus for the Copernicus Tokenizer, with the goal of providing broad token and pattern coverage rather than relying on a single text domain.

Dataset Overview

Property	Details
Total size	~12.3 GB
Domains	NLP, Code, Mathematics
Format	Parquet
Primary use	Tokenizer training / LM pretraining
Structure	Domain → Dataset source → Parquet shards
License	See individual source datasets

The dataset is organised into three primary domains:

Copernicus-Dataset/
├── nlp/
│   ├── fineweb-edu/
│   └── dolmino-mix-1124/
│
├── code/
│   ├── bigcode-starcoderdata/
│   ├── code_search_net/
│   └── stack-exchange-paired/
│
└── math/
    ├── MathInstruct/
    ├── finemath-4plus/
    ├── open-web-math/
    └── proof-pile-2-fixed/

Why combine different domains?

A tokenizer learns the statistical structure of the data it sees during training.

If the training corpus is dominated by general natural-language text, the resulting vocabulary will naturally favour the patterns, words, and symbols that occur in that domain.

That can lead to less efficient representations for:

• Source code
• Mathematical notation
• LaTeX
• Scientific terminology
• Structured technical text

The Copernicus dataset therefore combines natural language, code, and mathematics so that tokenizer training sees a much broader distribution of symbols, words, syntax, and domain-specific patterns.

The objective is not simply to make the dataset larger. It is to make the training distribution more diverse and useful.

Included Sources

Natural Language

FineWeb-Edu

Educational and high-quality web text used to provide broad natural-language coverage.

DOLMiNo Mix

A diverse mixture of high-quality text sources used to increase coverage beyond a single web corpus.

---

Code

BigCode / StarCoderData

Large-scale source-code data providing programming-language syntax, identifiers, operators, comments, and common software-development patterns.

CodeSearchNet

Code and natural-language pairs covering multiple programming languages and software-development terminology.

StackExchange-Paired

Technical question-and-answer content containing programming discussions, explanations, code fragments, and technical terminology.

---

Mathematics

MathInstruct

Instruction-oriented mathematical data containing mathematical problems, reasoning, and solutions.

FineMath-4+

Mathematical web content curated for high-quality mathematical text and notation.

OpenWebMath

Large-scale mathematical web text containing mathematical language and notation.

Proof-Pile-2

Mathematical and formal reasoning content containing mathematical expressions, proofs, and technical terminology.

Dataset Format

The dataset is stored in Parquet format and organised by domain and source.

This makes the dataset suitable for both:

1. Local training — download the required shards and process them locally.
2. Streaming training — load the dataset through the Hugging Face ecosystem and stream the data without downloading the complete corpus first.

Example:

from datasets import load_dataset

dataset = load_dataset(
    "Nj-1111/Copernicus-Dataset",
    streaming=True
)

For local workflows, individual Parquet shards can also be downloaded and processed directly with tools such as Hugging Face datasets, PyArrow, or pandas.

Intended Uses

This dataset can be used for:

• BPE / WordPiece / Unigram tokenizer training
• Custom tokenizer experimentation
• Small language-model pretraining
• Domain-general language-model experiments
• Code and mathematics language modelling
• Tokenization benchmarking
• NLP research and experimentation

The dataset is particularly useful when training a tokenizer that needs to cover general language + programming + mathematical text rather than optimising for a single domain.

Copernicus Tokenizer

This dataset was used to build the Copernicus Tokenizer, a 56K-token byte-level BPE tokenizer.

Tokenizer:

https://huggingface.co/Nj-1111/Copernicus-Tokenizer

The tokenizer was designed around the same objective as this dataset: broad coverage across natural language, code, mathematics, and technical text.

Reproducibility

The dataset is organised so that individual sources remain identifiable rather than being hidden inside one monolithic file.

This allows users to:

• Train on the complete collection
• Select individual domains
• Select individual source datasets
• Construct their own mixtures
• Experiment with different domain ratios
• Reproduce tokenizer-training experiments

Licensing

This repository is a curated collection of data originating from multiple datasets.

The licensing and usage terms of the original source datasets may differ.

Before using the data for commercial or redistributed applications, check the license and usage terms of each original dataset.

This repository does not override the licenses of the original sources.

Citation

If you use the Copernicus Dataset in your work, please reference this repository:

Copernicus Dataset
Nj-1111
Hugging Face Dataset Hub

Open Source

The dataset is publicly available so that others can experiment with tokenizer training and small language models without having to independently assemble the same collection of datasets.

Dataset: https://huggingface.co/datasets/Nj-1111/Copernicus-Dataset

Tokenizer: https://huggingface.co/Nj-1111/Copernicus-Tokenizer