pretty_name: Tiny Base 300M English Pretraining Tokens
license: other
language:
- en
tags:
- pretraining
- tokenized
- english
- reproducibility
size_categories:
- 1B<n<10B
Tiny Base 300M English Pretraining Tokens
This repository provides reproducible packed-token corpora for the Tiny Base 300M, 2,048-context base-model experiments. It contains uint16 token IDs, not raw source documents.
Available corpus versions
- Root-level files are the 50M-token pilot: 50,000,000 train tokens plus 1,000,000-token validation and test splits.
pretrain-312m-2k-1b/is the production corpus: 1,000,000,000 train tokens plus 1,000,000-token validation and test splits. It is the corpus used by the ongoing 312M-parameter training run.
Each version includes train.bin, validation.bin, test.bin, tokenizer.json, manifest.json, and settings.json. The binary files are little-endian uint16 streams. Use the matching tokenizer and verify SHA-256 values in the manifest before training.
Provenance and split policy
Both versions use 70% FineWeb-Edu, 25% English Wikipedia, and 5% TinyStories. Their manifests pin the exact Hugging Face dataset revisions and FineWeb-Edu configuration. Documents are deterministically split by SHA-256 of normalized text modulo 1,000: 0–9 test, 10–19 validation, and 20–999 train.
Terms and limitations
This derived token dataset remains subject to upstream terms: FineWeb-Edu (ODC-By and Common Crawl terms), English Wikipedia (CC BY-SA 3.0/GFDL), and TinyStories (CDLA-Sharing-1.0). Users are responsible for applicable attribution and share-alike obligations.
The corpus is English-first pretraining material, not a factual benchmark, safety filter, or curated knowledge base. It can retain bias, inaccuracies, and problematic material from public sources.