PiotrSty/polish-technical-mega-dynaword

Polish technical mega DynaWord — repaired v0.3.0

Polski korpus techniki i nauk ścisłych: przemysł, CNC, automatyka, informatyka, matematyka, fizyka i chemia. To teksty dokumentów, nie zbiór instrukcji do chatbota ani zatwierdzone procedury obsługi maszyn.

What changed

Chunks of one article/book now share one split. Shared PDF checksums, alternate source identities, ZPE page IDs and overlapping AGH book memberships are joined transitively. Wikibooks selection is an explicit review of book titles, not keyword deletion inside chapters. Out-of-scope content, historical technical scans and verified damaged article 103741 are preserved in quarantine with reasons. Text is not rewritten or completed by a language model. Samples cover source families rather than just the first alphabetic source.

Baseline revision: 66ea025cb1b47bdc08df2e4bed10c091a62a4f8c. All 22,331 original unique records remain accounted for: 15,809 retained, 6,522 quarantined. The previous revision remains available in repository history and under pre-repair-0.3.0.

Split	Text chunks	Document/book groups	Proxy tokens (cl100k_base)
train	13724	1874	25422631
valid	940	133	1822166
test	1145	108	4813267
quarantine	6522	196	8410264

Files and reproducibility

data/*.parquet keeps the eight-column DynaWord schema. The explicit default config excludes quarantine. provenance/records.jsonl maps every retained/quarantined record to a document group, content hashes, aliases, exclusion reasons and quality flags. provenance/source_metadata.jsonl preserves original attribution. processing/ contains the builder, explicit curation policy, packer and validators; rebuilding requires the original local source layers/metadata referenced by the builder, which are not all standalone Hub datasets. The published validation report checks the delivered artifacts without downloading original PDFs.

slayer_v6/ contains freshly regenerated little-endian uint16 native Slayer tokens, EOS 0 after each chunk and uint64 end offsets; see its manifest for exact token counts and tokenizer SHA256. manifest.json counts are proxy tokens, not native Slayer token counts.

Limits and training use

Group-disjoint splitting prevents the identified document/chunk leakage, not every near-duplicate, semantic overlap or pretraining contamination. All source text has not received expert factual/OCR review; remaining Unicode quality flags are explicit. Historical text is not current operating guidance. Licence metadata is retained, not independently cleared. Observe original attribution/share-alike terms.

Do not compare old and new validation curves as if the holdout were unchanged. Do not reuse old dataloader offsets with these regenerated packs. A model already trained on old held-out material does not regain an unseen test set through repartitioning.