SlayerLab/polish-dynaword-pretrainning-mix

Polish DynaWord Pretraining Mix

A Polish-language text corpus for causal language-model pretraining, derived exclusively from Polish DynaWord. It contains normalized, filtered and deduplicated documents with deterministic training, validation and test partitions, plus document-linked selection maps for four curriculum stages.

7.153B retained tokens · 3.955B training-target exposures · 1.718B unique selected targets

The four curriculum stages use sequence lengths of 512, 1,024, 2,048 and 2,048 tokens.

Dataset configurations

The default configuration contains the complete retained text pool: 2,994,054 documents. The four stage configurations contain ordered sequence references into that pool. Every configuration has train, validation and test splits.

Configuration	Content	Location
default	Full documents and record-level provenance, attribution and rights metadata	data/{split}/*.parquet
A512	Stage A sequence references	mixtures/A512/{split}/*.parquet
B1024	Stage B sequence references	mixtures/B1024/{split}/*.parquet
C2048	Stage C sequence references	mixtures/C2048/{split}/*.parquet
D2048	Stage D sequence references	mixtures/D2048/{split}/*.parquet

Load the dataset with Hugging Face Datasets:

from datasets import load_dataset

documents = load_dataset(
    "SlayerLab/polish-dynaword-pretrainning-mix", name="default", split="train", streaming=True
)
stage_a = load_dataset(
    "SlayerLab/polish-dynaword-pretrainning-mix", name="A512", split="train", streaming=True
)

The full text pool has 2,900,189 training, 47,097 validation and 46,768 test documents. Loading default selects this pool; the 3.955B training-target exposure schedule is represented by the A–D sequence maps.

The Parquet data comprises 96 files totaling 5,551,392,735 bytes (5.55 GB): 2,994,054 full-document rows and 6,204,978 ordered sequence references across the twelve stage/split maps. Normalized document text occupies 15,741,517,498 UTF-8 bytes before Parquet encoding.

Document fields

Fields	Meaning
id, text, split	Stable prepared document ID, complete normalized text and assigned split
source, source_id	DynaWord source and original record ID
upstream_dataset, upstream_revision, upstream_file, upstream_row	Pinned upstream dataset, source Parquet path and zero-based row
domain, quality_tier, quality_score	Preparation annotations
author, license	Original upstream author and rights-label values
license_expression, license_components, license_basis, license_urls	Normalized rights expression, component identifiers, declared basis and reference URLs
attribution_json, citations	Original source attribution sidecar entry serialized as JSON, and source citations
created, added	Upstream temporal metadata
group_id, text_sha256, privacy_status	Split-group identity, normalized-text hash and privacy-filter result

Sequence-reference fields

Each stage row contains sequence_index, doc_id, offset_in_document, token_count and target_count. Join doc_id to default.id in the corresponding split. sequence_index preserves the original stage order, including repeated exposures.

Offsets and lengths use the original preparation's token coordinates, including its document end marker; they are not character offsets. Each sequence has target_count = token_count - 1. The exported text is the complete normalized document; token totals below are measurements from the pinned preparation.

Source

All text comes from Polish DynaWord v0.2.6, pinned to revision 44a5f97e9d5a2a0ddc830c16d3890fbd60bbece8.

The complete source release comprises 27 Parquet shards, 3,108,752 records and 6,310,783,683 bytes. All shards were authenticated against their pinned sizes and SHA-256 digests and processed with the same preparation pipeline.

Polish DynaWord supplies encyclopedic, literary, parliamentary, legal, government, discussion, educational, news, satirical, scientific and mixed-reference text. Source identity and original author/license metadata are preserved in each exported document.

Source measurements: acquisition and corpus audit.

Corpus size

Quantity	Count
Source records processed	3,108,752
Normalized source-text tokens before filtering	8,168,386,362
Retained documents	2,994,054
Quarantined records	71,187
Exact duplicate records excluded	43,511
Retained stored tokens, including EOS	7,152,554,381
Training stored tokens, including EOS	6,873,000,723
Available training prediction targets	6,870,100,534
Validation prediction targets	139,273,551
Test prediction targets	140,186,242

Each stored document ends with one EOS token. Its initial token supplies conditioning context, so its prediction-target count is stored_token_count - 1. Prediction targets include EOS. Padding positions are ignored.

Curriculum stages

The mix consists of four generated training indexes over a shared token store.

Stage	Context length	Training-target exposures	Training sequences	Within-stage repeated targets
A — foundational	512	646,694,864	1,485,907	0
B — general/domain-rich	1,024	927,851,216	1,258,563	0
C — longer context	2,048	1,707,962,977	1,491,022	254,200,123
D — quality annealing	2,048	672,743,412	653,050	47,388,451
Total	—	3,955,252,469	4,888,542	301,588,574

Stage sizes reflect eligible source supply, domain/source/quality weights and per-stage repetition caps. Each stage uses the same held-out documents and validation/test target counts, with chunk boundaries determined by its context length.

Valid-target slot utilization is 85.00%, 72.00%, 55.93% and 50.30% for A–D respectively. Sequences remain document-local; shorter documents and final chunks use ignored padding.

The full selection parameters are recorded in the curriculum recipe. Measured sizes are recorded in the stage summary.

Mixture composition

The following percentages describe each domain's share of the generated stage's training-target exposures.

Domain	A512	B1024	C2048	D2048
Encyclopedic	39.84%	40.72%	36.20%	36.76%
Literature	21.45%	14.97%	16.25%	21.45%
Parliamentary	12.75%	14.81%	14.48%	9.19%
Legal	4.78%	9.26%	12.67%	6.13%
Government	6.37%	8.09%	8.74%	6.13%
Discussion	6.37%	5.55%	5.43%	6.13%
Educational	2.82%	1.95%	2.10%	5.11%
News	2.61%	1.82%	1.96%	4.93%
Satire	1.59%	1.85%	1.10%	1.53%
Scientific	0.92%	0.64%	0.69%	1.71%
Other	0.49%	0.34%	0.36%	0.94%

Percentages are rounded. Legal and parliamentary text together account for 22.84% of mix exposures, compared with 68.34% of retained stored tokens. The Other category is principally the mixed-genre NKJP pool.

Exact domain counts and proportions: mixture composition.

Unique coverage and repetition

Across all four stages, the mix selects 1,718,473,498 distinct prediction-target positions, covering 25.0138% of the 6,870,100,534 available training targets.

The 3,955,252,469 total exposures comprise those distinct targets plus 2,236,778,971 repeated exposures:

• 301,588,574 repeats within stages.
• 1,935,190,397 additional exposures from overlap between stages.

The shared training pool contains a further 5,151,627,036 targets outside this selection.

Domain	Available training targets	Selected unique targets	Unique coverage	Share of total exposures
Encyclopedic	1,367,149,151	722,317,553	52.8339%	37.95%
Literature	139,190,936	139,122,162	99.9506%	17.69%
Parliamentary	1,622,963,903	332,388,793	20.4804%	13.37%
Legal	3,077,757,854	267,757,339	8.6998%	9.47%
Government	79,345,519	77,722,123	97.9540%	7.76%
Discussion	520,621,181	116,097,591	22.2998%	5.73%
Educational	18,267,326	18,263,674	99.9800%	2.70%
News	16,870,964	16,870,563	99.9976%	2.54%
Satire	18,836,683	18,836,683	100.0000%	1.43%
Scientific	5,921,328	5,921,328	100.0000%	0.89%
Other	3,175,689	3,175,689	100.0000%	0.48%

Unique coverage is calculated by taking the union of document-local target intervals [offset + 1, offset + count) across all stage indexes, with training-split and document-boundary checks.

Measurements: exact coverage and domain coverage.

Preparation

Normalization and filtering

Text normalization applies Unicode NFC, consistent newlines and spaces, trailing-space cleanup, outer-whitespace removal and bounded blank lines while preserving Polish diacritics and paragraph boundaries.

Source-informed domain labels, format cues and heuristic quality scores assign records to high_general, high_specialized, supplementary or quarantine. Filtering covers language cues, encoding problems, extreme repetition, source-rights conditions and recognizable personal-data patterns.

Rights/privacy pattern checks cover whole documents. Language, format and repetition analysis uses the first and last 50,000 characters, with at most 12,000 sampled words for language cues. The preparation character ceiling is 5,000,000; token counts were measured for every processed record.

Biblioteka Nauki and TED procurement records are quarantined under the source-rights policy. The retained chemistry source is grouped at source level and assigned to a held-out partition.

Deduplication and splits

Exact normalized-text copies are grouped before final split assignment, and retained duplicate copies are stored once.

The partition seed is 20261103, with 0.02 validation and 0.02 test group-assignment fractions. Grouping follows documents and source-specific families, including NKJP parent documents, whole-source chemistry groups and cross-source parliamentary calendar days where available. All chunks inherit their parent document's split.

Each stage's validation and test indexes cover the same held-out documents. Training-specific weights, score thresholds and length filters apply to training selection.

Stage selection

Parameter	A512	B1024	C2048	D2048
Minimum document tokens, including EOS	33	65	129	65
Minimum heuristic quality score	0.60	0.65	0.65	0.75
Supplementary-tier multiplier	0.25	0.30	0.25	0.10
Maximum within-stage exposures	1	1	2	2

The maximum document length is 2,000,000 tokens for every stage. High-general and high-specialized quality multipliers are 1.0. EUR-Lex, Dziennik Ustaw, both wiki-talk pools and satire have a stricter one-exposure cap per stage.

Source multipliers operate within domains alongside available target counts and quality weights. Each pool exhausts its unique targets before a later exposure pass. Seeded chunk selection and interleaving preserve document boundaries. Adjacent chunks share one conditioning token, and repeated exposures are represented by index entries over the shared token store.

Document provenance

Each document retains its Polish DynaWord source name, original record ID, source Parquet path, zero-based row number and pinned dataset revision, together with the original author and license metadata.

Normalized text, domain and quality annotations, split assignment, group identity and exact-text hashes are exported alongside these fields. Original attribution sidecars and Wikipedia citations are joined by source and record ID. Original empty author fields remain unchanged; additional credits and issuing-institution details remain available in the source attribution metadata and notices.

Every curriculum reference points back to an exported document and its document-local position in the original preparation's token coordinates. Repeated selections retain the same document identity.

Verification

Full-corpus verification covered all 3,108,752 processed records and 2,994,054 retained documents, checking file identities, provenance, document boundaries and split separation.

All four stages passed complete validation/test traversal and resume replay checks.

Export verification compares all 2,994,054 documents and 6,204,978 sequence references against the prepared catalog, authenticated attribution database and original stage indexes. It checks complete text and metadata, row identities, sequence order, split separation, license counts, file hashes and byte accounting.

All five Hugging Face configurations and their fifteen splits load locally in offline streaming mode; the loader check exercises the first record and declared row counts in every split.

Results: preparation verification, exhaustive export verification and Hugging Face loader check.

Licenses and attribution

The retained corpus contains 26 original rights-label strings, normalized to 24 expressions with 14 distinct license/status components across 25 sources. Normalization preserves AND/OR semantics and jurisdiction-specific versions; the original license field remains unchanged.

The following document groups are mutually exclusive:

Declared terms or status	Documents
CC BY, including combinations of versions or jurisdiction ports	43,788
CC BY-SA, including combinations of versions or jurisdiction ports	1,991,824
Combined CC BY and CC BY-SA requirements	4
Licence Art Libre 1.3	1,467
EU reuse policy	229,427
Public-domain status declared upstream	4,382
Polish official-document status declared upstream	723,162
Total	2,994,054

The collection's selection, arrangement and original metadata are offered under CC BY-SA 4.0, to the extent those contributions are protected and licensable by their contributor. Constituent texts retain their own terms. Official-document and public-domain labels describe upstream legal bases, not grants of a CC0 license. Hugging Face metadata therefore identifies the package as license: other, with mixed source terms.

Computed accounting: license summary, machine-readable counts, source obligations, attribution notice and pinned upstream evidence.

Attribution, share-alike, edition, translation and quotation obligations follow the underlying material. Specific conditions include Sejm Information System terms for commercial services or products, EU reuse exceptions, historical Wikimedia/imported-content licenses and territorial limits of official-document status. The source-obligation register and unresolved obligations preserve these conditions; the accounting does not establish blanket legal clearance for every downstream use.

Data characteristics

Privacy filtering quarantines records with detected emails, labelled contact numbers, PESEL candidates and some residential-address patterns. These checks apply to document text; attribution and citation metadata retain upstream values. Heuristic detection can leave residual personal information and can flag institutional contacts. Exact-text deduplication can leave related editions, overlapping legal revisions, reprinted passages and shared templates across groups.

The corpus reflects its source registers: encyclopedic and formal prose, historical literature, political speech, community discussion and satire. These sources contain stylistic and representational biases, dated claims, harmful language and factual errors. Domain labels and quality scores are source-informed heuristics.

References

• Polish DynaWord v0.2.6
• Upstream corpus description
• Upstream notice
• Curriculum recipe
• Corpus audit
• Mixture composition
• Unique target coverage