VardanSharma13/spectrum-pretrain

Spectrum pretraining data (v3)

Built by Vardan Sharma (MSc AI, University of Edinburgh) for the Spectrum project with Vivek Iyer and Prof. Alexandra Birch.

Cleaned documents for pretraining Spectrum, an encoder-decoder LM over OmniSONAR sentence embeddings (OmniSONAR paper, Section 9). One document per row. Sentence splitting is done downstream, so there is no sentences column; n_sentences_rule is only a rough rule-based count.

Quick start

from datasets import load_dataset
ds = load_dataset("VardanSharma13/spectrum-pretrain", "hin", split="train")   # or "multilingual", "english"
ds[0]["text"]

Size

• Multilingual: 58 languages, 31,972,412 documents, 17,170,212,072 words, about 1.04B sentences (rule count).
• English: 21,677,915 documents, 12,231,931,595 words, about 0.66B sentences (rule count). FineWeb-Edu, Cosmopedia and Common Corpus. DCLM-Edu is not included: the project already holds it, and FineWeb-Edu pages that also appear in DCLM-Edu were removed (4,913,103 documents).
• Speech: VoxPopuli, 16 languages, 75,776 documents, 379,665 segments, 1,050.77 hours.

Columns (text)

column	meaning
id	unique id, <lang>-<shard>-<row>
orig_id	the document's id in our download of its source
text	the cleaned document
lang	ISO 639-3 code
source / subset	where it came from (fineweb2, sangraha, wikipedia, fineweb_edu, cosmopedia, common_corpus)
doc_type	web, wiki, synthetic, ...
url, domain	origin page (when known)
n_words, n_chars, n_sentences_rule	sizes
lid_prob	GlotLID document probability for lang
fragile_frac	share of lines removed as SONAR-fragile (URLs, references, IDs, number rows, menus)
wrong_lang_frac	share of lines removed as another language
quality_score	English only: FineWeb-Edu classifier score 0-5 (FineWeb-Edu and Common Corpus; NaN for Cosmopedia)
meta	JSON string with the source's own fields

How it was cleaned

Each step follows a published pipeline (OmniSONAR/Spectrum, LCM, FineWeb/FineWeb-2, DCLM, MADLAD-400, Sangraha).

1. Document filters: Unicode normalisation, script check, GlotLID language ID, Gopher/FineWeb repetition and quality rules, emoji and mojibake rules, PII removal (emails, phone numbers, IPs).
2. Blocklist and spam: UT1 adult/gambling domain blocklist, plus a spam-keyword density rule.
3. SONAR-fragile lines (LCM paper): lines that are URLs, citation/reference lines, ID tokens, number or table rows, and runs of short menu labels are removed. A document losing more than 30% of its text this way is dropped.
4. Line-level language ID (MADLAD-400): lines of 40 or more characters confidently (p >= 0.8) in another language are removed. A document losing more than 30% is dropped.
5. Deduplication: exact (normalised text) and MinHash near-duplicates (128 permutations, 5-word shingles, Jaccard 0.8), across all sources of a language, Wikipedia kept first. For English: exact across all sources, MinHash on Common Corpus (FineWeb-Edu is already deduplicated per dump, and the FineWeb paper found cross-dump dedup hurts).
6. Boilerplate lines: lines repeated across many documents of a language (cookie banners, shop footers) are removed.
7. Domain cap: no single domain above 2% of a language (Wikipedia exempt).
8. Decontamination: documents sharing n-grams with FLORES, XStoryCloze, Belebele, SIB-200 (and HellaSwag for English) are removed.
9. Final length check: at least 3 sentences and 30 words (100 characters for languages without spaces).
10. English quality: Common Corpus documents scoring below 2 on the FineWeb-Edu classifier are dropped (the DCLM-Edu threshold).

Per language

flagged raw and flagged clean are the share of a 20,000-document sample that any quality check flags, before and after cleaning.

lang	docs	words	sentences (rule)	near dups removed	flagged raw	flagged clean
arb	814,305	372,023,056	20,512,626	758	36.2%	35.2%
asm	188,105	81,893,233	7,268,709	46,210	12.3%	7.2%
ben	583,716	222,876,543	20,049,353	72	5.9%	4.5%
brx	2,858	649,648	36,670	2,858	30.8%	27.0%
bul	504,408	251,052,313	18,742,916	449	8.2%	6.5%
cat	637,678	253,668,023	13,321,122	3,142	28.7%	26.6%
ces	636,674	254,830,323	22,229,435	238	8.0%	3.5%
cmn	1,178,203	1,420,425,272	43,860,671	79	4.7%	2.3%
dan	538,180	234,313,324	17,925,634	1,069	12.9%	4.2%
deu	1,127,348	475,065,846	39,480,121	173	12.1%	6.8%
ell	795,486	331,427,767	20,148,103	2,649	24.8%	23.3%
est	326,662	164,764,310	16,151,577	1,177	8.5%	4.7%
fas	742,001	449,771,322	22,499,590	1,382	21.9%	19.8%
fin	420,345	153,589,402	16,027,211	1,151	12.5%	2.6%
fra	1,112,377	532,663,584	34,515,968	363	16.1%	13.6%
gle	169,974	92,362,144	5,705,413	13,234	29.4%	22.4%
gom	9,594	2,989,082	256,832	9,719	35.6%	4.4%
guj	451,426	184,217,844	13,109,817	32,069	10.2%	9.6%
heb	451,031	219,460,838	16,057,964	1,654	6.0%	4.3%
hin	759,810	312,124,801	18,422,392	14,861	8.2%	5.1%
hrv	255,296	211,161,896	16,689,152	5,786	17.7%	16.4%
hun	565,734	264,460,187	21,844,021	169	12.9%	9.8%
ind	709,425	335,922,395	24,505,043	3,336	20.0%	11.3%
ita	990,153	468,564,599	27,906,469	1,049	31.2%	28.8%
jpn	959,153	1,183,394,810	41,991,359	23	4.7%	1.6%
kan	490,659	120,013,602	11,761,103	35,649	7.1%	5.8%
kor	639,607	219,643,739	22,096,823	1,876	6.6%	1.2%
lit	364,159	189,928,371	17,559,339	1,194	9.1%	7.4%
lvs	311,616	156,397,289	13,993,802	2,923	11.6%	8.4%
mai	12,302	4,979,380	369,940	12,485	38.2%	5.8%
mal	587,379	103,757,237	11,442,694	27,365	5.5%	3.9%
mar	520,699	164,313,130	13,555,333	13,231	6.0%	5.2%
mlt	131,533	59,903,447	4,280,634	13,165	23.9%	17.9%
mni	763	310,037	18,473	763	50.1%	40.2%
nep	679,695	194,183,730	13,690,582	393	6.9%	6.8%
nld	817,679	325,649,188	25,403,108	1,095	7.8%	2.7%
ori	423,295	118,290,740	10,785,068	226	6.4%	4.9%
pan	379,983	178,849,928	9,950,836	25,895	11.2%	10.5%
pol	882,714	337,625,776	30,264,702	213	7.1%	5.6%
por	1,031,417	471,137,594	29,963,299	325	18.9%	17.0%
ron	697,307	338,884,188	22,274,317	1,913	19.0%	17.3%
rus	956,589	622,568,237	55,520,856	229	8.2%	7.2%
san	160,611	69,642,670	11,929,975	26,460	26.9%	15.4%
sat	98	32,283	2,330	98	9.1%	5.1%
slk	526,035	203,340,133	18,122,115	2,593	9.0%	4.9%
slv	385,879	194,377,597	15,643,101	2,109	12.3%	7.6%
snd	219,066	129,274,039	6,423,180	68,223	21.3%	20.7%
spa	1,367,246	688,941,269	36,503,071	302	32.1%	30.4%
srp	297,835	172,817,107	12,464,023	5,743	16.2%	15.5%
swe	589,717	233,156,505	17,686,369	1,567	11.7%	2.7%
swh	253,706	101,741,446	6,069,163	12,424	25.2%	22.4%
tam	572,376	155,290,435	15,242,492	4,064	10.5%	9.4%
tel	516,044	127,374,686	13,514,011	150	6.7%	4.2%
tha	626,734	1,645,400,877	11,584,745	1,348	10.7%	9.3%
tur	770,765	274,295,833	24,140,722	311	15.4%	11.2%
ukr	677,753	253,283,619	21,523,470	1,051	11.5%	8.9%
urd	566,671	327,794,944	12,828,345	117	38.1%	32.1%
vie	584,538	513,340,454	23,513,092	1,200	13.2%	10.9%
eng	21,677,915	12,231,931,595	659,419,906	709,945

Speech: VoxPopuli

European Parliament speeches from facebook/voxpopuli (CC0), as ordered documents: a document is an unbroken run of at least 3 consecutive segments of one speech. Every segment was checked: duration 1-30 s, silence, clipping, speaking rate, and Whisper large-v3-turbo transcription compared with the VoxPopuli text (character error rate must be <= 0.5). A document is split where a segment is dropped; sentence order is never broken. Audio is 16 kHz mono 16-bit FLAC inside the parquet. FLEURS is excluded because it is a Spectrum evaluation set.

language	docs	segments	hours	segments dropped	median CER
en	20,583	99,097	275.82	6.19%	0.0182
de	13,271	67,300	165.49	9.02%	0.0419
fr	9,569	48,656	132.75	2.91%	0.0159
es	7,420	41,113	119.82	3.55%	0.0104
pl	5,073	24,659	68.73	3.33%	0.0095
ro	4,092	20,634	64.3	3.61%	0.0221
it	2,901	15,862	53.77	14.89%	0.0373
hu	2,768	14,429	41.89	5.47%	0.0154
cs	3,026	14,905	41.12	2.71%	0.0189
nl	2,481	11,257	24.84	7.57%	0.0353
sk	1,822	8,598	23.93	3.25%	0.0203
hr	1,048	4,998	15.42	3.93%	0.0273
fi	1,156	5,410	15.03	2.63%	0.0159
sl	383	1,857	5.42	20.38%	0.0513
et	124	605	1.58	1.93%	0.0321
lt	59	285	0.86	12.97%	0.0758

Segments: seg_id, doc_id, position, text, flac, sample_rate, duration_s, asr_cer. Docs: doc_id, orig_doc_id, speech, speaker, seg_ids, texts, durations_s, n_segments, total_s.

Checksums

SHA256SUMS.text and SHA256SUMS.speech list a sha256 for every file.

Licences

Each source keeps its own licence: FineWeb-2 (ODC-By), Sangraha (CC-BY-4.0), Wikipedia (CC-BY-SA-4.0), FineWeb-Edu (ODC-By), Cosmopedia (Apache-2.0), Common Corpus (open licences per document), VoxPopuli (CC0). Use within those terms.