docs1

The dataset of high-quality, diverse, factual documents for in-context use.

docs1 is AlphaQualia's base corpus of documents for in-context use. It contains factual, high-quality English documents of 1K–32K tokens, sorted into 31 length buckets of 1,024 MiniCPM5-2B tokens each. It is the shared starting point for every downstream dataset (questions, reconstructions, tool tasks, teacher generations). Those datasets refer to documents by id and never copy them.

Documents	2,889,645
Tokens (MiniCPM5-2B)	34.81B
Length range	1,024 – 32,768 tokens, 31 buckets
Docs per bucket	25,560 (longest) to 181,521 (shortest)
Sources	23 public Hub datasets
Exact duplicates removed	79,462
Version	v1.1 (stage 1: curation + light cleanup), built 2026-10-14

Usage

from datasets import load_dataset

# Everything (streams; about 34.81B tokens of text)
ds = load_dataset("alpha-qualia/docs1", split="train", streaming=True)

# One length bucket: 8,192–9,215 tokens
ds = load_dataset("alpha-qualia/docs1", data_files="data/b07_8192-9215/*.parquet", split="train")

# A length range, by exact token count
ds = ds.filter(lambda r: 8500 <= r["length_tokens_minicpm"] <= 9000)

Short documents are far more common than long ones (see the figure), so for a uniform length distribution, sample the same number of documents from every bucket folder. Every bucket has at least 25,560.

Files are stored as data/b{NN}_{low}-{high}/part-*.parquet, so a curriculum (short first, then long) can read whole buckets. Rows are shuffled within each bucket.

Schema

Field	Type	Description
id	string (UUID)	Stable document ID: UUIDv5 of source repo : source row id (plus the character span for segments). A rebuild gives the same IDs.
created_at	timestamp (UTC)	When the document was added to docs1.
source_type	string	The Hub dataset it came from (e.g. HuggingFaceFW/finepdfs-edu), or original for documents we write (later stages).
metadata	string (JSON)	Source fields such as url, title, license, repo, path, plus source_file. Segments add parent_id, segment, char_span.
content	string	The document text, unchanged except for segment cuts.
length_tokens_minicpm	int32	Exact token count with the MiniCPM5-2B tokenizer (no special tokens).
tags	list[string]	Labels added by later stages, e.g. uhq (top-quality subset), synthetic, augmented. Empty in v1. Filter with "uhq" in r["tags"].

Composition

Source	Hub dataset	Category	Docs	Tokens	Mean length	Segmented	License
finepdfs-edu	HuggingFaceFW/finepdfs-edu	PDFs	499,065	7.34B	14,704	16%	ODC-By 1.0
fineweb-edu	HuggingFaceFW/fineweb-edu	Web	268,881	3.79B	14,095	14%	ODC-By 1.0
gutenberg	common-pile/project_gutenberg_filtered	Books	192,444	3.11B	16,174	92%	open (per doc, Common Pile)
caselaw	common-pile/caselaw_access_project_filtered	Law	237,961	2.93B	12,323	6%	open (per doc, Common Pile)
usgpo	common-pile/usgpo_filtered	Government	195,675	2.78B	14,191	41%	open (per doc, Common Pile)
pes2o	common-pile/peS2o_filtered	Science	218,957	2.35B	10,719	1%	open (per doc, Common Pile)
youtube	common-pile/youtube_filtered	Transcripts	205,885	2.23B	10,853	5%	open (per doc, Common Pile)
finewiki	HuggingFaceFW/finewiki	Wikipedia	187,693	1.87B	9,970	6%	CC BY-SA 4.0 / GFDL
stack-edu	common-pile/stackv2_edu_filtered	Code	176,001	1.81B	10,281	11%	open (per doc, Common Pile)
doab	common-pile/doab_filtered	Books	151,778	1.76B	11,565	43%	open (per doc, Common Pile)
pubmed	common-pile/pubmed_filtered	Science	123,865	1.75B	14,145	5%	open (per doc, Common Pile)
ar5iv	marin-community/ar5iv-no-problem-markdown	Science	126,461	1.21B	9,566	1%	arXiv (per paper)
stackexchange	common-pile/stackexchange_filtered	Q&A	95,889	570.6M	5,950	0%	open (per doc, Common Pile)
rules	Andwwy/rules	Agent files	78,808	558.7M	7,089	4%	per repository (unverified)
regulations	common-pile/regulations_filtered	Government	42,438	297.0M	6,999	12%	open (per doc, Common Pile)
uk-hansard	common-pile/uk_hansard_filtered	Transcripts	10,255	159.3M	15,534	87%	open (per doc, Common Pile)
skillmd	FayeZC/SkillMD-138K	Agent files	26,832	104.1M	3,880	0%	per file (card: CC BY 4.0)
pressbooks	common-pile/pressbooks_filtered	Textbooks	23,301	99.3M	4,260	1%	open (per doc, Common Pile)
libretexts	common-pile/libretexts_filtered	Textbooks	17,987	65.0M	3,612	1%	open (per doc, Common Pile)
news	common-pile/news_filtered	News	6,404	11.3M	1,764	0%	open (per doc, Common Pile)
oercommons	common-pile/oercommons_filtered	Textbooks	2,043	9.0M	4,402	3%	open (per doc, Common Pile)
peps	common-pile/python_enhancement_proposals_filtered	Tech docs	597	2.6M	4,342	0%	open (per doc, Common Pile)
public-domain-review	common-pile/public_domain_review_filtered	Essays	425	1.1M	2,554	0%	open (per doc, Common Pile)

Per-bucket counts

Bucket (tokens)	Docs	Tokens	Largest sources
1,024–2,047	181,521	267.2M	finepdfs-edu 11%, youtube 7%, pes2o 7%
2,048–3,071	175,166	440.3M	finepdfs-edu 11%, youtube 7%, pes2o 7%
3,072–4,095	168,259	596.8M	finepdfs-edu 11%, youtube 7%, pes2o 7%
4,096–5,119	163,511	748.4M	finepdfs-edu 12%, youtube 7%, pes2o 7%
5,120–6,143	155,657	871.9M	finepdfs-edu 12%, youtube 8%, pes2o 8%
6,144–7,167	150,080	994.8M	finepdfs-edu 13%, pes2o 8%, finewiki 8%
7,168–8,191	144,972	1.11B	finepdfs-edu 13%, pes2o 8%, finewiki 8%
8,192–9,215	138,440	1.20B	finepdfs-edu 14%, pes2o 9%, finewiki 9%
9,216–10,239	132,471	1.29B	finepdfs-edu 15%, pes2o 9%, finewiki 9%
10,240–11,263	128,087	1.37B	finepdfs-edu 15%, pes2o 9%, finewiki 9%
11,264–12,287	122,528	1.44B	finepdfs-edu 16%, pes2o 10%, caselaw 9%
12,288–13,311	115,851	1.48B	finepdfs-edu 17%, pes2o 10%, caselaw 10%
13,312–14,335	108,186	1.49B	finepdfs-edu 18%, pes2o 11%, caselaw 11%
14,336–15,359	100,785	1.49B	finepdfs-edu 19%, pes2o 12%, caselaw 11%
15,360–16,383	94,119	1.49B	finepdfs-edu 21%, pes2o 13%, caselaw 11%
16,384–17,407	86,953	1.47B	finepdfs-edu 22%, pes2o 11%, fineweb-edu 11%
17,408–18,431	80,582	1.44B	finepdfs-edu 24%, fineweb-edu 11%, caselaw 11%
18,432–19,455	74,208	1.40B	finepdfs-edu 25%, fineweb-edu 12%, caselaw 10%
19,456–20,479	68,718	1.37B	finepdfs-edu 26%, fineweb-edu 12%, caselaw 9%
20,480–21,503	63,270	1.33B	finepdfs-edu 27%, fineweb-edu 13%, gutenberg 9%
21,504–22,527	57,848	1.27B	finepdfs-edu 27%, fineweb-edu 13%, gutenberg 10%
22,528–23,551	52,651	1.21B	finepdfs-edu 27%, fineweb-edu 14%, gutenberg 12%
23,552–24,575	47,671	1.15B	finepdfs-edu 27%, fineweb-edu 14%, gutenberg 12%
24,576–25,599	43,812	1.10B	finepdfs-edu 27%, fineweb-edu 14%, gutenberg 13%
25,600–26,623	40,642	1.06B	finepdfs-edu 27%, fineweb-edu 14%, gutenberg 14%
26,624–27,647	37,858	1.03B	finepdfs-edu 27%, gutenberg 15%, fineweb-edu 14%
27,648–28,671	35,840	1.01B	finepdfs-edu 27%, gutenberg 16%, fineweb-edu 14%
28,672–29,695	33,405	974.7M	finepdfs-edu 27%, gutenberg 17%, fineweb-edu 14%
29,696–30,719	31,342	946.7M	finepdfs-edu 26%, gutenberg 18%, fineweb-edu 14%
30,720–31,743	29,652	925.9M	finepdfs-edu 27%, gutenberg 18%, fineweb-edu 14%
31,744–32,768	25,560	823.7M	finepdfs-edu 28%, gutenberg 19%, fineweb-edu 14%

Construction

1. Sources. Datasets that are already curated and trusted: FineWeb-Edu, FinePDFs-Edu, FineWiki, ar5iv, the openly licensed Common Pile v0.1 subsets (papers, books, textbooks, Q&A, government, law, transcripts, educational code), and agent instruction files (Andwwy/rules, SkillMD).
2. Filters. Source-level quality filters only: FinePDFs-Edu not truncated and without near-duplicates (minhash_cluster_size == 1); Stack-Edu code with edu score ≥ 3; agent files mostly Latin script and with no credential patterns.
3. Length. Every document is tokenized with the MiniCPM5-2B tokenizer. Documents of 1,024–32,768 tokens are kept whole. Longer documents (mostly books, theses and long reports) are cut into at most 6 consecutive segments at paragraph breaks. Each segment is sized for a random length bucket that still needs documents.
4. Balance. Each source collects up to a fixed number of documents per bucket (6,000; FinePDFs-Edu 12,000), so no single source can dominate a bucket. The merge removes exact duplicates and keeps every remaining document; buckets are not truncated, so balancing by length happens at load time.
5. Determinism. Sources, files and rows are chosen with fixed seeds. Building again from the same Hub revisions gives the same IDs and nearly the same selection.

Cleaning (v1.1)

Conservative, rule-based, chosen by reading flagged documents from a 3% stratified sample; rules with real false positives were narrowed to the sources where they only caught junk.

Step	Documents
dropped: repeated	2,864
dropped: table_dump	2,843
dropped: symbol_soup	2,254
dropped: non_english	1,187
dropped: nav_menu	82
dropped: out_of_range_after_fix	35
repaired: nul	13,553
repaired: mojibake	2,236
moved to another bucket after repair	601

non_english: mostly non-Latin text; symbol_soup: under 60% letters, digits and spaces (prose sources); repeated: over half the text in repeated lines (PDFs and US government sources); table_dump: number tables of one or two words per line; nav_menu: scraped site navigation. nul: NUL bytes removed; mojibake: wrong-encoding text such as ’ repaired with ftfy. Repaired documents keep their id and list the repairs in metadata.cleaned. Earlier versions stay available in the repository history.

Limitations

• Long buckets are less diverse. Natural documents over ~16K tokens are rare outside PDFs, papers and books, so those categories dominate the long end. Later stages add synthetic long documents.
• Not decontaminated against any benchmark. Check overlap before using docs1 near an evaluation set.
• Novelty varies. Wikipedia and famous texts are likely known to the base model; papers, PDFs and agent files much less so.
• Mixed licenses. The repository is marked CC BY-NC-ND 4.0, but each document also keeps its source license (see metadata.license where present; e.g. Wikipedia is CC BY-SA). Check the source license before reuse.
• English only, by source selection (no language filter on Common Pile subsets beyond their own).

Roadmap

• Stage 2, original documents: newly scraped material (recent docs, repositories, papers) and model-written documents aimed at specific length buckets (source_type = "original").
• Stage 3, augmented documents: model-made variants of docs1 documents, from small factual edits to style changes. Each records metadata.parent_id.

Reproduce

# Stage 1: one CPU job per source group (8 in parallel)
uv run gpu --provider aws --region us-east-1 sagemaker submit datasets/docs1/collect.sh --instance ml.c6i.32xlarge --name aq-docs1-<group>
# Stage 2: merge, card, publish
uv run gpu --provider aws --region us-east-1 sagemaker submit datasets/docs1/merge.sh --instance ml.r6i.16xlarge --name aq-docs1-merge

Code: datasets/docs1/ in the AlphaQualia repository.