DATA_CARD.md

Data card: LLM Speedrun token streams

Summary

This repository contains two shuffled, pre-tokenized text streams and the tokenizer used to create them. The artifacts are intended for research and educational language-model training.

Data sources and processing

• fineweb-edu-10BT.shuffle.bin comes from the sample/10BT configuration of HuggingFaceFW/fineweb-edu, using its text field.
• smoltalk.shuffle.bin comes from the data/all configuration of HuggingFaceTB/smoltalk, using its messages field. Each conversation was serialized as <ROLE>content</ROLE> messages inside <DOCUMENT>...</DOCUMENT>.
• Documents from each source were shuffled with Python's random.seed(42) before tokenization.
• tokenizer_50M.bpe was trained on the first 40,000,000 bytes of the shuffled FineWeb-Edu text and the first 10,000,000 bytes of the shuffled SmolTalk text. It is a JSON-serialized byte-level BPE with 32,520 defined vocabulary entries. IDs 32512–32519 are the document, system, user, and assistant opening/closing tags; the associated model vocabulary space is 32,768.
• The token streams were written as headerless, little-endian unsigned 16-bit integers. They contain 9,440,023,113 FineWeb-Edu tokens and 875,269,408 SmolTalk tokens.

The repository provides token streams rather than row-oriented examples and has no predefined train, validation, or test split.

Intended use

The data is intended for language-model pretraining, instruction-data mixing, and reproducible educational experiments using the matching tokenizer implementation. It is not prepared for direct viewing with the Hugging Face dataset viewer.

Licensing

This is a mixed-source derived dataset, so the repository is marked other. FineWeb-Edu is released under ODC-By 1.0 and is also subject to Common Crawl's terms of use. SmolTalk combines newly generated Apache-2.0 subsets with existing public datasets governed by their respective licenses. Users are responsible for reviewing and complying with the upstream dataset cards and terms.

Limitations and risks

No additional safety, privacy, bias, or quality audit was performed after obtaining the upstream datasets. FineWeb-Edu is derived from web data and may contain harmful, biased, toxic, copyrighted, or personal information despite upstream filtering. SmolTalk includes synthetic and repackaged public instruction data and may contain generation errors or inherit limitations from its component datasets. The binary streams do not retain document-level metadata, source identifiers, or original dataset splits.