mistobaan/fineweb10B_bytes

FineWeb 10B Bytes

This repository contains training shards for byte-level language model pretraining.

The dataset format is the same format used by openai/parameter-golf, the OpenAI Model Craft Challenge repository for training compact language models and evaluating them on FineWeb in bits per byte. In that repository, evaluation is described as tokenizer-agnostic and based on compression performance on the FineWeb validation set. oai_citation:0‡GitHub

Origin

These byte shards were generated using the data conversion approach from openai/parameter-golf Pull Request #705, authored by GitHub user seanward. That PR is titled “Byte-Level Tokenizer-Free Transformer” and explicitly includes a conversion script named convert_to_bytes.py, described there as “Data conversion (sp1024 → raw bytes)”. oai_citation:1‡GitHub

Contents

This repository stores shard files such as:

• fineweb_train_000000.bin
• fineweb_train_000001.bin
• fineweb_train_000002.bin

and so on.

Dataset format

The .bin shards follow the same binary training-data convention used for byte-level experiments in parameter-golf.

At a high level:

• data is represented as raw UTF-8 bytes
• the byte vocabulary size is 256
• shards are intended for training tokenizer-free / byte-level models
• the data layout is meant for efficient streaming during pretraining

The associated PR #705 describes the model as operating directly on raw UTF-8 bytes with vocab=256, and states that it uses raw byte input without BPE or SentencePiece. oai_citation:2‡GitHub

Provenance

Source data is derived from FineWeb preprocessing workflows associated with byte-level training experiments for parameter-golf.

This repository republishes the resulting training shards only. It does not bundle the training code itself; for the original training setup, conversion logic, and experiment context, see:

• openai/parameter-golf
• PR #705 by seanward (“Byte-Level Tokenizer-Free Transformer”) oai_citation:3‡GitHub

Intended use

This dataset is intended for:

• byte-level language model pretraining
• tokenizer-free training experiments
• reproducing or adapting parameter-golf-style data pipelines
• benchmarking compact models on byte-level objectives