fengluoqiuwu/owt-seg512

OWT GPT-2 sentence segments (fixed pad 512)

Tokenized sentence/paragraph segments of OpenWebText (Skylion007/openwebtext, revision main), encoded with GPT-2 BPE plus three extra special tokens.

This repository is a derived dataset. It does not rediscover or replace the original text corpus. Every input_ids sequence comes from documents in OpenWebText.

• Packaging / conversion code & this card: MIT (see LICENSE)
• Underlying web text: inherited from OpenWebText / Skylion007/openwebtext. Use of the token sequences remains subject to the original corpus terms and to the licenses of the crawled pages.

中文摘要

本仓库是 OpenWebText 的派生数据：用 GPT-2 BPE 按句段切分，定长 pad 到 512。validation 对应本地训练里的 dev。内容与 bdelf 本地缓存 owt_owt-seg512_c661b6d8c3f996d0（fingerprint c661b6d8c3f996d0）逐 token 一致。仓库协议为 MIT；文本来源仍归 OpenWebText。

What's in a row

column	type	meaning
input_ids	list[int32]	tokens without pad, length = length. Always starts with BOS and ends with EOS.
length	int32	len(input_ids), in [128, 512]

To restore the exact bdelf training row (right-padded to 512):

PAD_ID = 50257
row = example["input_ids"] + [PAD_ID] * (512 - example["length"])

Pad tokens are not stored, to keep the Hub copy smaller. They are a deterministic fill with PAD_ID.

Splits

OpenWebText on Hub only ships a train split. Document-level holdout is applied on parquet row order of Skylion007/openwebtext:

Hub split	bdelf name	#chunks	source
train	train	20,618,358	remaining documents
validation	dev	2,524	1,024 documents
test	test	2,721	1,024 documents

Holdout: shuffle indices 0 .. N-1 with holdout_seed=42, take first 1,024 rows → test documents, next 1,024 → dev documents. Chunk counts differ because documents vary in length.

Tokenization

item	value
Base tokenizer	Hugging Face gpt2 (GPT-2 BPE, Fast)
Encode	add_special_tokens=False
Added specials (in this order)	<|pad|>=50257, <|bos|>=50258, <|eos|>=50259
Layout vocab_size	50260
Unused original EOT	GPT-2 <|endoftext|> = 50256 is not used as BOS/EOS here

Each segment is wrapped as [BOS] + content + [EOS]. Segments with wrapped length ≤ 128 are dropped (min_chunk_len=128).

Segmentation (must match training)

Documents are not concatenated. For each document:

1. GPT-2 Fast encode with offset_mapping (byte BPE; punctuation is matched on UTF-8 text, then mapped to token boundaries).
2. Content cap d - 2 = 510 (room for BOS/EOS), process_d=512.
3. If remaining tokens exceed the cap, look back min(256, d/8)=64 tokens and cut on the best delimiter: paragraph blank lines (P0) → single newline (P1) → sentence punctuation (P2). If none, hard-cut at the cap and continue from that index (no tokens dropped).
4. Right-pad to 512 with PAD_ID for training storage; this Hub export strips that pad.

Block shuffle after cutting: shuffle_seed=42, block size 65,536 rows (bdelf shuffle_mode: block).

Fingerprint of the local cache this export was taken from: c661b6d8c3f996d0.

Load

from datasets import load_dataset

ds = load_dataset("fengluoqiuwu/owt-seg512")
print(ds)
# DatasetDict with train / validation / test

ex = ds["train"][0]
print(len(ex["input_ids"]), ex["length"])  # equal

Streaming:

ds = load_dataset("fengluoqiuwu/owt-seg512", split="train", streaming=True)

Relation to the sibling dataset

fengluoqiuwu/owt-bucket uses the same documents and holdout, but process_d=2048 and bucket pads {256, 512, 1024, 2048}. A 512-length slice from this repo is not the same distribution as the 512-bucket of owt-bucket (different cut points). Do not mix them in one training step if you need the bdelf length-curriculum protocol.

Source / citation

Primary source: Skylion007/openwebtext.

@misc{Gokaslan2019OpenWeb,
  title={OpenWebText Corpus},
  author={Aaron Gokaslan and Vanya Cohen},
  year={2019},
  howpublished={\url{https://skylion007.github.io/OpenWebTextCorpus/}}
}

Please also cite this derived segmentation if you use these exact splits and cuts.

License

MIT for the files in this Hub repository (conversion, metadata, token-id tables). The original OpenWebText documents remain a scrape of public web pages; this repo does not grant additional rights to that underlying text.