quintic/fineweb-scaled-gpt2

FineWeb Scaled GPT-2 Prefixes

This repository contains nested 2B, 4B, 8B, and hero-scale token prefixes for controlled language-model scaling experiments. The binary shards use the llm.c GPT-2 v1 format and are directly consumable by the GPT TPU Speedrun trainer.

Dataset structure

Each folder is independently usable after its manifest.json is present:

folder	validation tokens	training tokens
2B/	100,000,000	1,900,000,000
4B/	100,000,000	3,900,000,000
8B/	100,000,000	7,900,000,000
hero/	100,000,000	74,900,000,000

The variants are exact nested prefixes: the validation shard and early training shards have identical SHA-256 values across folders. Repeated Hub paths are expected to deduplicate at the content-storage layer.

Every .bin file has a 1,024-byte header of 256 little-endian signed int32 values. Header slots 0, 1, and 2 contain magic 20240520, version 1, and the token count. The remaining payload is exactly 100,000,000 little-endian uint16 GPT-2 token IDs.

Source and preparation

The source is HuggingFaceFW/fineweb_100BT-shuffled at revision ee8552966e3d6a5fee2f317f2ae0b342be03d998. Its card describes a global document shuffle with seed 42. Each variant includes source.json with the immutable revision plus the byte length and LFS SHA-256 of all 100 source Parquet files.

To maintain temporal separation from the Fresh10 diagnostic corpus, source rows are retained only when their crawl/source date is strictly before 2024-04-01; missing or invalid dates are rejected. The 40 normalized Fresh10 source URLs and canonical-text hashes are excluded defensively. Fresh10 raw artifact hashes are also compared opportunistically with UTF-8 source-row bytes, but are not treated as normalized-content hashes. exclusions.json contains the exact policy, and each manifest records observed exclusion counts at that prefix boundary.

Retained text is re-tokenized—not inferred from the source token_count column—with GPT-2 tiktoken==0.11.0. Token 50256 is inserted before every document. The first exact 100M tokens are validation and all remaining shards are training. Ordinary training-shard boundaries may split a document. At the validation boundary only, the remainder of the crossing document is discarded so validation and training are document-disjoint; the manifest records the discarded-token count and a hash-safe document identifier.

The preparation code uses one verified source Parquet at a time, bounded PyArrow batches, streaming output writes, exact row-level checkpoints, atomic shard installation, and SHA-256 verification. The generated manifests are the authoritative per-file integrity inventory.

The publication includes the exact frozen preparation sources at provenance/fineweb_builder.py and provenance/prepare_fineweb.py. Their SHA-256 values are respectively 26c61bc921af290e6beb28596feb2c50cac5b15a56a2f3adf921682317f6f109 and 3a676241de10c3ac7cf36ed19ccbd1c0e419bb90de960d4e14be51a1f225bd5c. The same frozen sources are independently retrievable from immutable Git commit c6acab32cea6e48260d139be1774b3e3286d7afd. Every production variant pins source-inventory SHA-256 02ddc6361cc2f8a3d23b0d8b823c7eb7e2b1663ad3d0eff63e83b373456fc12b, exclusion-policy SHA-256 ab25cabd0781b1046b7ad7b281b4147ff6e27d36977f4e842b8c92573399ad77, and preparation-core SHA-256 4bbdcb76da837276f6f337b805d37a74e3272b476e01fd198f416097abe19241.

Intended use

These variants are intended for controlled pretraining and IsoFLOP/scaling-law experiments where data prefixes, tokenizer, split, and binary layout must be held constant. They are not intended as a benchmark test set or as a source of factual ground truth.

Limitations and responsible use

FineWeb derives from Common Crawl. Despite upstream filtering, it can contain harmful, biased, inaccurate, copyrighted, or personally identifying material. The temporal cutoff and Fresh10 exclusions address evaluation leakage, not all possible duplication or contamination. Users remain responsible for suitable content controls, legal review, and downstream safety evaluation.

License and attribution

FineWeb is distributed under ODC-By 1.0 and remains subject to the Common Crawl Terms of Use and underlying source terms. The GPT TPU Speedrun preparation code is Apache-2.0; that code license does not replace the corpus license. Please cite and attribute the upstream FineWeb dataset and paper and the shuffled source's requested SmolData citation:

@misc{niklaus2026smoldata,
  title        = {SmolData},
  author       = {Joel Niklaus and Hynek Kydl{\'\i}{\v{c}}ek},
  year         = {2026},
  publisher    = {Hugging Face},
  journal      = {Hugging Face repository},
  howpublished = {\url{https://huggingface.co/collections/HuggingFaceFW/smol-data}}
}

Retain this preparation provenance when redistributing shards.