File size: 1,835 Bytes
b78c6d5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | ---
license: odc-by
language:
- en
size_categories:
- 10B<n<100B
tags:
- byte-level
- language-modeling
- pretraining
---
# FineWeb-Edu as raw bytes (sample-10BT)
[FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu)
(`sample-10BT` config) prepared for **byte-level language modeling**, in the
spirit of [nanoGPT](https://github.com/karpathy/nanoGPT)'s `train.bin` /
`val.bin` — except here there is no tokenizer at all: **one uint8 per UTF-8
byte of text**. ~45 GB total.
Made for training tokenizer-free models (byte-level Transformers,
[H-Nets](https://arxiv.org/abs/2507.07955), etc.) and comparing them against
BPE baselines on identical data.
## Format
Flat binary shards of `uint8`, directly memmap-able:
```
train_00000.bin ... train_000NN.bin (500 MB each)
val_00000.bin
meta.json
```
- Bytes `0..253`: raw UTF-8 bytes of the text
- Byte `254`: `<|doc|>` document separator (0xFE never occurs in valid UTF-8).
Every document **starts** with one.
- Byte `255`: reserved, never occurs
- Every 1000th document went to the `val` split; exact document and byte
counts are in `meta.json`.
## Usage
```python
import numpy as np
shard = np.memmap("train_00000.bin", dtype=np.uint8, mode="r")
# random crop as training example
x = shard[1_000_000 : 1_000_000 + 8192]
# decode back to text
docs = bytes(shard[:100_000]).split(b"\xfe")
print(docs[1].decode("utf-8", errors="replace")[:200])
```
## Provenance
Produced by streaming `HuggingFaceFW/fineweb-edu` `sample-10BT` in order,
encoding each `text` field as UTF-8, and concatenating with separators. No
filtering, shuffling, or other processing was applied (occurrences of bytes
0xFE/0xFF, impossible in valid UTF-8, would be stripped — in practice there
are none). License follows the source dataset (ODC-By, subject to
CommonCrawl ToU).
|