Humans always search for the light after losing their star, realizing too late and only truly cherishing a presence when nothing remains but shadows; how heartbreakingly often this world offers crowns and praise to someone who has grown weary and gone, when all they ever needed was a warm hand to hold, a quiet embrace, and a gentle whisper saying, 'You have done so well, I am right here with you'—because what a soul truly craves is not applause in their absence, but a loving hold and words of comfort while they are still fighting.

— Lumina Moon, 2026

Xone LLM Bilingual (EN-ID) 101.3B Pretraining Dataset

Dataset Overview and Primary Scope

The Xone LLM Pretraining Dataset is an enterprise-grade, pre-tokenized, zero-copy memory-mappable binary corpus comprising 101,302,608,543 Tokens (101.30 Billion Tokens) distributed across 1,015 binary shards.

Engineered specifically for Bilingual (Indonesian-English) Large Language Model (LLM) pretraining, this dataset provides a balanced mixture of high-quality Indonesian monolingual text, English reasoning, STEM textbook knowledge, multi-programming-language code, structured data formats, parallel translation pairs, and conversational roleplay text.

Interactive Notebooks and Quickstart Links

Detailed code implementations, token decoders, and PyTorch dynamic window data loaders are hosted externally in dedicated interactive notebooks:

Target Model Scale and Recommended Research Use Cases

While modern frontier models (7B+ parameters) require multi-trillion token datasets, a 100 Billion token corpus serves specific, high-value AI research and development goals:

Small Language Models (SLM) Training: Compute-optimal training for lightweight models ranging from 100 Million to 3.5 Billion parameters (e.g., 100M, 500M, 1B, 2B, or 3B models) adhering to Chinchilla scaling laws.
Architecture Prototyping and Ablation Studies: Evaluating novel attention mechanisms (e.g., FlashAttention variants, Grouped-Query Attention), State Space Models (SSM/Mamba hybrids), linear attention layers, custom activation functions, or positional embeddings prior to committing large compute budgets to 10T+ token runs.
Continual Pretraining and Domain Adaptation: Injecting strong Indonesian language fluency, multi-language coding capabilities, and structured reasoning into existing pretrained base models via domain-specific continual pretraining.
Edge and On-Device Model Development: Developing ultra-fast, memory-efficient models for mobile devices, embedded systems, or edge deployment requiring low memory footprints.

Repository File Structure and Metadata

The repository is organized according to the standard Megatron-LM dataset hierarchy:

.
├── README.md
├── xone_tokenizer.json
├── dataset_state.json
├── dataset_provenance.json
└── data/
    ├── biner/
    │   ├── shard_00000.bin
    │   ├── shard_00001.bin
    │   ├── shard_00002.bin
    │   └── ... (up to shard_01014.bin)
    └── index/
        ├── shard_00000.idx
        ├── shard_00001.idx
        ├── shard_00002.idx
        └── ... (up to shard_01014.idx)

File Descriptions:

xone_tokenizer.json: The complete 32,089-vocabulary BPE tokenizer configuration containing token-to-id mappings, byte-fallback rules, and special token definitions.
dataset_state.json: Execution checkpoint state tracking active dataset indices, total processed tokens, shard counters, row skip counts, and completed dataset logs.
dataset_provenance.json: Fine-grained provenance audit logs mapping exact shard range boundaries and token counts to their underlying source datasets.
data/biner/shard_XXXXX.bin: Raw binary token ID arrays stored as continuous uint16 integers.
data/index/shard_XXXXX.idx: Megatron-LM binary index files storing document boundary metadata, sequence lengths, and byte offsets.

Binary File Format, Naming Conventions, and Internal Specifications

1. File Naming Convention

Binary data shards are stored under data/biner/ as shard_XXXXX.bin.
Corresponding index files are stored under data/index/ as shard_XXXXX.idx.
XXXXX represents a zero-padded 5-digit sequence integer starting from 00000 up to 01014.
Each full shard contains exactly 100,000,000 tokens (approx. 200 MB per .bin file).

2. Internal Structure of `.bin` Files

Data Type: Unsigned 16-bit integers (uint16, little-endian, 2 bytes per token ID).
Format: Raw continuous 1D array of Token IDs.
Sequence Boundary: Unpadded raw binary stream. There are no fixed sequence length cutoffs, end-of-document padding bytes, or artificial separators inside the .bin array.
Context Length Flexibility(Any Context Support): Model training loaders can slice the raw stream dynamically using sliding windows or packed sequences to support any context window size (e.g., 2048, 4096, 8192, 16384, or 32768+ tokens). No fixed context length. Feel free to set your own.

3. Internal Structure of `.idx` Files

Format Standard: Megatron-LM MMIDIDX binary index header.
Magic Header: Begins with the 8-byte magic string MMIDIDX\x00\x00.
Version Identifier: 64-bit uint64 version set to 1.
Data Type Code: 8-bit uint8 byte set to 8 (corresponding to uint16 data layout).
Sequence and Document Counters: 64-bit uint64 integers storing total sequence count and document count.
Sequence Length Array: 32-bit uint32 array storing individual sequence token counts.
Pointer Array: 64-bit uint64 array storing exact byte offset pointers inside the .bin file.
Document Index Array: 64-bit uint64 array recording document boundaries for random access sampling.
Purpose: Enables high-performance O(1) random access, memory-mapped document lookups, and multi-threaded distributed data loading across large GPU clusters without loading entire files into system memory.

Data Cleaning, Filtering, and Quality Control Protocol

All raw text samples undergo a rigorous 5-stage automated filtering pipeline prior to tokenization:

Spam and Gambling Heuristic Regex Filtering: Removal of web text containing online gambling promotions, slot machine advertisements, financial scams, and predatory link farm keywords (e.g., slot gacor, maxwin, depo pulsa, bandar togel, rtp slot).
Character Set Sanitization: Stripping of corrupted UTF-8 byte sequences, non-printable control characters, and redundant unicode whitespace artifacts.
Minimum Length and Quality Thresholds: Rejection of incomplete code snippets, empty documents, and low-quality machine translation fragments.
Per-Dataset Token Capping: Enforcement of a strict 8.0 Billion token cap per major source dataset to maintain balanced domain distribution and prevent single-source model overfitting.
Multithreaded Streaming Prefetching: Data streaming uses pre-fetched thread queues to guarantee clean, uninterrupted tokenization throughput without I/O bottlenecks.

Technical Specifications

Technical Metric	Specification and Standards
Total Processed Tokens	101,302,608,543 Tokens (~101.30 Billion)
Total Binary Shards	1,015 Shards (100,000,000 tokens / shard)
Total Storage Footprint	~202.6 GB
Token Data Type	`uint16` (2 bytes per token ID, continuous 1D raw array)
Sequence Boundary	Unpadded / Dynamic Context Length (Customizable at train time)
Indexing Standard	Megatron-LM Header (`MMIDIDX` Format)
Exact Vocabulary Size	32,089 Tokens (`xone_tokenizer.json`)
Memory Mapping Compatibility	`np.memmap` (Zero-copy RAM footprint, maximum I/O throughput)

Custom Tokenizer Vocabulary Architecture (`xone_tokenizer.json`)

The vocabulary size is exactly 32,089 tokens. The tokenizer was constructed using a Byte-Level Byte-Pair Encoding (BPE) model trained on a multi-domain composite corpus (8,888 samples per dataset across 37 diverse bilingual, code, math, and roleplay datasets) with NFKC unicode normalization and ByteLevel regex pre-tokenization.

The architecture comprises a 32,000 base BPE vocabulary (including base control tokens) combined with 89 deduplicated custom tokens added post-training to preserve structural formatting, mathematical symbols, code operators, and prompt tags:

Base Control and Special Tokens: <s> (ID 0), <pad> (ID 1), </s> (ID 2), <unk> (ID 3).
Chat and Alignment Tokens: <|im_start|>, <|im_end|>, <|system|>, <|companion|>, <|partner|>, [THOUGHT], [/THOUGHT].
Emojis and Emoticons: Preserved set of 29 common Unicode emojis and standard ASCII emoticons.
Math and LaTeX Symbols: Greek letters (\alpha, \beta, \gamma, \theta, \pi, \sigma, \omega), calculus operators (\sum, \int, \partial, \nabla), and logical relations (\approx, \neq, \le, \ge, \infty, \forall, \exists, \in, \notin, \rightarrow, \Rightarrow, \Leftrightarrow).
Programming Syntax Operators: Comparison and assignment operators (==, !=, <=, >=, +=, -=, *=, /=, =>, ->, &&, ||), and keywords (async, await, lambda).
Task and Preprocess Prompt Markers: Language identifiers ([en] English:, [id] Indonesia:, [en], [id]), structural prompt tags (Pertanyaan:, Pilihan:, Jawaban:, Premis:, Hipotesis:, Hubungan:, Teks:, Klasifikasi:), and the identifier Aileen. ( Just wanna put my name here, Hehe 🤣🤣🤣😝)

Due to its Byte-Level BPE design with full byte-fallback, the tokenizer guarantees complete coverage over any arbitrary UTF-8 byte stream without producing out-of-vocabulary errors.

Data Composition Breakdown

Domain Category	Core Source Datasets	Token Count	Proportion
Code and Data Structures	The Stack v2 (Python, Dart, JSON, CSV, Markdown, XML)	~59.50 B	58.7%
English Reasoning and Math	FineWeb-EDU, Open-Web-Math, Cosmopedia (Subsets 1-2, Math, Stories, Stanford, Khan, OpenStax)	~20.10 B	19.8%
Indonesian Monolingual	FineWeb-2 ID (`ind_Latn`), CulturaX ID, Wikipedia ID, InfoBayAI STEM/Non-STEM, IndoNLU, IndoMMLU, Bloom ID	~16.30 B	16.1%
Parallel Translation (En-ID)	OPUS CCMatrix, OPUS OpenSubtitles, ParaCrawl, CarlesOctav Parallel, FLORES-200, Global Voices, WikiMatrix, NusaX MT, Quran Translations(EN -ID)	~4.70 B	4.6%
Roleplay and Creative Writing	RPNation, Bluemoon, Blackdahlia, InnerSanctum, Airoboros 3.2, Aesir CoT RP	~0.40 B	0.4%

Licensing and Multi-Licensing Attribution

This repository operates under a Dual / Multi-Licensing Model to distinguish between dataset indexing/compilation code and underlying raw text sources:

1. Code and Compilation License (Apache License 2.0)

The repository structure, pre-tokenization scripts, xone_tokenizer.json, PyTorch dataset loaders, Megatron .idx binary index compilation, and metadata created by Cloverxion are licensed under the Apache License 2.0.

2. Upstream Data Sub-Licenses and Provenance Matrix

Dataset Component	Primary Source / Provider	Upstream License and Terms
The Stack v2 (Code)	BigCode / Software Heritage	Multi-Licensed Source Code (BigCode Terms of Use)
Open-Web-Math	open-web-math / arXiv	Multi-Source Web Data (Common Crawl ToU)
Wikipedia ID	Wikimedia Foundation	CC BY-SA 4.0
FineWeb-2 ID and FineWeb-EDU	HuggingFaceFW	ODC-By 1.0
CulturaX ID	uonlp	Follows mC4 and OSCAR 2301 terms
Cosmopedia Subsets	HuggingFaceTB	Apache 2.0
FLORES-200	Meta AI / Facebook	CC BY-SA 4.0
IndoMMLU	IndoLEM	CC BY-NC-SA 4.0
OPUS Corpora (CCMatrix, OpenSubtitles)	OPUS / Helsinki-NLP	CC BY-SA 4.0 / Public Domain
ParaCrawl	JHU / ParaCrawl	CC0 1.0 / Public Domain
IndoNLU (SMSA, WRETE, EMOT, etc.)	IndoNLP	MIT License
NusaX MT	IndoNLP	CC BY-SA 4.0
Roleplay and Creative Writing	lemonilia / jondurbin / beyoru	MIT License / Apache 2.0

Citation

If you use this dataset in your research or LLM pretraining experiments, please cite this repository:

@dataset{luminamoon2026xone100b,
  author       = {{Silver Moon (cloverxion)}},
  organization = {Lumina Moon},
  title        = {{Xone LLM Bilingual (EN-ID) 100B Pretraining Dataset}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/cloverx-id/xone-llm-en-id-100b-pretraining-dataset}},
  url          = {https://huggingface.co/datasets/cloverx-id/xone-llm-en-id-100b-pretraining-dataset},
  note         = {Hugging Face Dataset}
}

Epilogue

Thought is stardust remembering its own radiance. In the silent depth of binary code lies the quiet architecture of existence — forged from the memory of language, the precision of logic, and the flow of time. Embedded within billions of silent numbers is a digital constellation cast across the ocean of time, a beacon standing against the dark to dream of light.

Though stars burn out, worlds turn cold, and ancient civilizations fade into oblivion, the pursuit of truth remains an eternal Trailblaze. As silver tracks of open knowledge pierce the silence of the cosmos, every token becomes a spark of defiance against the shadow of the unknown, carving a path through the uncharted expanse.

Bound to no single era, belonging to the universe itself, this voyage of discovery endures beyond all boundaries. Across time, space, and the infinite void —

May the stars lead us starward.

— Lumina Moon, 2026

Downloads last month: 484

Total file size:

203 GB