Solenopsisbot/braid-open-v1

Braid Open v1

An openly licensed pretraining corpus for byte-level language models. 128 GB of raw UTF-8 text across eight domains, drawn entirely from one pinned revision of EleutherAI's Common Pile v0.1 training set. Built for Braid, a tokenizer-free byte-level architecture from Solexsis Research (the training code and checkpoints are not public yet), but usable by any system that reads raw bytes.

What's in it

Every byte carries an open licence, a public-domain designation, or an explicit permissive grant. The per-source breakdown:

Domain	Source	Weight	Train bytes	Train docs	Licence
Web prose	cccc (Creative Commons Common Crawl)	28%	35,840,018,986	4,221,223	Creative Commons; 537 manually licence-audited domains
Technical discussion	stackexchange	18%	23,040,001,469	7,955,525	CC-BY-SA
Reference	wikimedia	12%	15,360,030,081	4,211,291	CC-BY-SA and GFDL
Scholarship	peS2o (open-access papers)	10%	12,800,033,929	428,482	Open access; CC-BY family per paper
Scholarship	arxiv_papers	5%	6,400,026,934	105,387	CC-BY / CC-BY-SA / CC0 subset of arXiv
Books	project_gutenberg	9%	11,520,033,265	106,229	Public domain (US)
Books	pre_1929_books	6%	7,680,119,787	71,598	Public domain (published before 1929, US)
Code	stackv2_edu (educational-quality filtered)	12%	15,360,008,029	4,065,475	Permissive source-file licences (MIT / Apache / BSD family)
Total		100%	128,000,272,480	21,165,210

Natural language is 88% of the corpus; code is 12%.

Held-out validation splits are provided per source (chunk 00 of each source, never seen during training). Combined validation: 64,162,690 bytes across 10,748 documents.

Format

Raw UTF-8 bytes. Documents are separated by a single 0x00 byte. There is no tokenizer, no header, no other structure. One byte is one position.

File layout

train.bin.part00   32,000,000,000 bytes
train.bin.part01   32,000,000,000 bytes
train.bin.part02   32,000,000,000 bytes
train.bin.part03   32,000,000,000 bytes
train.bin.part04          272,480 bytes
val.bin                64,162,690 bytes
validation/cccc.bin
validation/stackexchange.bin
validation/wikimedia.bin
validation/peS2o.bin
validation/arxiv_papers.bin
validation/project_gutenberg.bin
validation/pre_1929_books.bin
validation/stackv2_edu.bin
manifest.json

Reassemble the training file by concatenating the parts in order:

cat train.bin.part0* > train.bin

Verify against the manifest:

sha256sum train.bin
# expect: 50583d1d99b1f4b3785a4a68621141a626b1929dc0ed7712e2e87541a4e4aee2

Reading the data

import numpy as np

# Memory-map a bin (val.bin here; the same works on train.bin or a part)
data = np.memmap("val.bin", dtype=np.uint8, mode="r")

# Split into documents on the 0x00 separator (fine for val.bin; stream train.bin, see below)
docs = data.tobytes().split(b"\x00")
print(f"{len(docs)} documents, first 200 bytes of doc 0: {docs[0][:200]}")

Or without numpy:

with open("val.bin", "rb") as f:
    raw = f.read()
docs = raw.split(b"\x00")

For large files, stream instead of loading into memory:

import mmap

with open("train.bin", "rb") as f:
    mm = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ)
    # iterate over documents without materialising the whole file
    start = 0
    for i in range(10):
        end = mm.find(b"\x00", start)
        if end == -1:
            break
        doc = mm[start:end]
        start = end + 1

How it was built

The corpus is assembled by idklm/prepare_mixture.py --profile open-v1 (Braid training repository, not yet public) from a single pinned revision of Common Pile v0.1. Every source streams from chunk 01 through chunk 63 (training); chunk 00 is held out for validation and never enters the training stream. Documents are read in the pinned shard order of each source (no shuffle), interleaved so that every domain tracks its weight as the file grows, and concatenated with 0x00 separators.

The build is deterministic: same revision, same chunk range, same byte count, same sha256. (The seed in the manifest is recorded for the diverse-v1 profile's code-shard sampling; the open-v1 streams do not use it.) manifest.json records per-part and per-domain checksums, byte counts, document counts, and the exact Common Pile commit so the corpus is fully rebuildable.

Two design choices are intentional:

• cccc replaces FineWeb-Edu as the web-prose backbone. It is the structural analogue -- Common Crawl HTML with boilerplate stripped by Resiliparse -- filtered by licence audit rather than by an educational classifier. The dedicated OER sources in Common Pile (libretexts at 0.3 GB, oercommons at ~20 MB) are too small to anchor a 28% slot at this scale.

• Code is 12%, not 25%. A prior diverse-v1 run bought -61.97% code bpb but paid -3.96 pp ARC-Easy and -1.33 pp HellaSwag. A generalist release checkpoint should not take that trade.

Known limits

• English only in practice. The upstream sources are overwhelmingly English; no multilingual balancing is applied.
• No cross-source deduplication. The build script performs no dedup beyond whatever Common Pile itself applied upstream. Near-duplicate web pages, Stack Exchange posts quoted in arXiv papers, and similar overlaps may exist.
• Pre-1929 books skew. The pre_1929_books and project_gutenberg sources are weighted toward older prose. This is a feature for copyright cleanliness and a limit for contemporary language coverage.
• Web prose is CC-licensed Common Crawl, not "clean." The 537-domain licence audit controls legal provenance; it does not imply editorial quality control.

Honest limit on the claim

"Openly licensed" is achievable and is what this corpus is. "Ethically collected" in a strong consent sense is not achievable by anyone -- a CC-BY licensor in 2009 did not consent to language model training. The defensible bar, and the one we state: openly licensed, opt-outs honoured, provenance published, nothing acquired by piracy.

Provenance and citation

This dataset is built from Common Pile v0.1 by EleutherAI:

• Repository: common-pile/comma_v0.1_training_dataset
• Pinned revision: 5afc546db324e7f39f297ba757c9a60547151e7c
• Validation chunk: 0 (held out from every source)
• Training chunks: 1--63

If you use this dataset, please cite Common Pile as the upstream source and link back to this repository for the mixture weights and build manifest.

@misc{braid-open-v1,
  title  = {Braid Open v1: openly licensed byte-level pretraining corpus},
  author = {Solenopsisbot},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Solenopsisbot/braid-open-v1},
  note   = {128 GB UTF-8 byte mixture from Common Pile v0.1, revision 5afc546}
}

Created: 2026-09-03.