Squash: collapse LFS history to reclaim private storage

.gitattributes

@@ -0,0 +1,60 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.avro filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.lz4 filter=lfs diff=lfs merge=lfs -text
+*.mds filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+# Audio files - uncompressed
+*.pcm filter=lfs diff=lfs merge=lfs -text
+*.sam filter=lfs diff=lfs merge=lfs -text
+*.raw filter=lfs diff=lfs merge=lfs -text
+# Audio files - compressed
+*.aac filter=lfs diff=lfs merge=lfs -text
+*.flac filter=lfs diff=lfs merge=lfs -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text
+*.ogg filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+# Image files - uncompressed
+*.bmp filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.tiff filter=lfs diff=lfs merge=lfs -text
+# Image files - compressed
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text
+*.webp filter=lfs diff=lfs merge=lfs -text
+# Video files - compressed
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.webm filter=lfs diff=lfs merge=lfs -text

BLOGPOST.md

@@ -0,0 +1,96 @@
+# Licensed to Train: a 100% openly-licensed Polish LLM stack
+
+*Draft — Slayer Labs. Doubles as the skeleton for a short tech report.*
+
+## TL;DR
+
+- We built **Polish DynaWord**, a Polish pre-training corpus where **every single token is openly licensed for model training**, with a documented legal basis per source.
+- **5.64B tokens · 2.38M documents · 6 sources** — larger than Danish Dynaword (4.8B), the reference edition in the family.
+- We trained a **dedicated 32k Polish BPE** that encodes Polish **1.75× tighter than Bielik's tokenizer** and **1.58× tighter than Llama-3** — at a quarter of Llama-3's vocabulary.
+- We are training a small LM **from scratch on this corpus** with a modern stack (modded-nanogpt: Muon, RoPE, RMSNorm, QK-norm, SwiGLU). The result is a model whose **entire provenance — data, tokenizer, and training — is legally defensible end-to-end.**
+- Everything is released: corpus, tokenizer, and reproducible build scripts.
+
+---
+
+## 1. The problem: clean data is the hard part
+
+Most "open" pre-training datasets ride on Common Crawl, whose underlying content carries ambiguous rights. That ambiguity is not academic: projects have been pulled or never released over exactly this (the Dynaword paper, [arXiv:2508.02271](https://arxiv.org/abs/2508.02271), cites a shut-down music generator, a withdrawn Danish encoder, and the never-released Nordic Pile).
+
+For Polish specifically, there was no openly-licensed, traceable, continuously-maintained corpus in the [Dynaword](https://huggingface.co/datasets/danish-foundation-models/danish-dynaword) family. Danish, Norwegian, Swedish, Icelandic — yes. Polish — no. That gap is the project.
+
+The interesting work is **not** downloading text. It is the licensing, deduplication, decontamination, and provenance — putting each source in the right place with the right paper trail. That is what takes time; the bytes are commodity.
+
+## 2. The corpus
+
+We follow the Dynaword guidelines literally:
+
+1. **Open & traceable licensing.** Not "this is public domain" but *why* — statutory basis or upstream license, documented per source.
+2. **Reproducibility.** A single script rebuilds the corpus from sources.
+3. **Documented.** A datasheet per source.
+4. **Extensibility.** Versioned; new sources by contribution.
+
+**Sources (v0.1):**
+
+| source | license | documents | tokens |
+|---|---|---:|---:|
+| Legal & official (EU law + parliamentary) | CC-BY-4.0 / official documents (art. 4) | 567k | 4.0B |
+| Encyclopedic & reference (Wikipedia + Wikisource) | CC-BY-SA-3.0 | 1.80M | 1.5B |
+| Literature & books (public-domain collections) | CC-BY-SA-4.0 / public domain | 6k | 0.1B |
+| **total** | | **2.38M** | **5.64B** |
+
+Only **human-authored** text — no synthetic, machine-translated, or auto-transcribed data. Gates are deliberately minimal (drop short docs, non-Polish, exact duplicates, OCR garble); heavy quality filtering and mix-weighting belong downstream, in training, not in the released corpus.
+
+**What we excluded, and why** (transparency matters as much as inclusion):
+
+- **Movie/TV subtitles** — derivative of copyrighted dialogue; mostly unlicensed uploads. This is the exact lesson Danish Gigaword learned with OpenSubtitles. Out.
+- **A large cultural-heritage aggregate** — mixed per-record rights (public-domain *and* rights-reserved in one bucket). Needs per-record rights filtering before any inclusion.
+- **A "novels" corpus** — labelled permissive, but likely contains in-copyright contemporary works. Pending title-level verification.
+
+The data itself comes from an open redistribution of these upstream sources; the redistributor is credited, and each upstream license/attribution is preserved.
+
+### Sidebar: the download was its own little puzzle
+
+The source host caps throughput at **~1.5 MB/s per IP** (confirmed identical across three machines — it is per-IP policing, not our link). Single-stream TCP collapsed under packet loss; parallel connections plateaued at the per-IP ceiling. The fix was simply **more IPs**: we sharded the corpus — and byte-range-split the 8.9 GB long-pole file — across three machines in parallel, reassembling on the target box. A reminder that the ceiling is rarely where you first think it is.
+
+## 3. The tokenizer: a Polish BPE that beats Bielik
+
+nanoGPT-style stacks default to GPT-2's byte-level BPE — which is terrible on Polish. So we trained a **32k byte-level Polish BPE** on a domain-balanced ~1 GB sample (capping the 71%-legal skew so the tokenizer doesn't overfit to legalese). Training took **51 seconds**.
+
+Then we measured **fertility** (tokens per word) on held-out Polish, against the obvious baselines:
+
+![tokenizer fertility](polish_tokenizer_fertility.png)
+
+| tokenizer | vocab | tok/word | vs ours |
+|---|---:|---:|---:|
+| **Polish-32k (ours)** | 32,768 | **1.74** | 1.00× |
+| Llama-3 | 128,000 | 2.74 | 1.58× worse |
+| Bielik-11B-v3 | 32,000 | 3.04 | **1.75× worse** |
+| GPT-2 | 50,257 | 3.66 | 2.10× worse |
+
+The headline: **our 32k Polish BPE encodes Polish 1.75× more efficiently than Bielik's** — at the *same* vocabulary size. And Bielik is *worse than Llama-3*, despite Llama having a 4× larger vocab.
+
+Why? **Bielik-11B-v3 is built on Mistral and inherited Mistral's tokenizer** — a general multilingual 32k vocab, never trained on Polish. It never got a Polish-native tokenizer. A dedicated Polish BPE sees the same text roughly twice as densely.
+
+Practically this means: ~2× more text per context window, and the same FLOPs train on ~2× more Polish. (Caveat for the rigorous: the held-out sample is in-domain Wikipedia; all four tokenizers see identical text, and the 1.75× gap is too large to be a domain artifact, but a fully out-of-domain re-test strengthens the published number.)
+
+## 4. The model: clean provenance, end to end
+
+A model is only as "open" as its *entire* lineage. Continually pre-training someone else's checkpoint inherits that checkpoint's data-provenance questions. So we train **from scratch** on the clean corpus.
+
+We use **modded-nanogpt** (Keller Jordan's speedrun stack — Muon optimizer, RoPE, RMSNorm, QK-norm, SwiGLU, fp8): a modern 124M-scale architecture, already validated, far more FLOP-efficient than vanilla GPT-2. We swap in our Polish BPE and retokenize the corpus into training shards.
+
+Target: a 124M (and later ~350M) Polish LM trained entirely on openly-licensed data, in ~one night on a single H100.
+
+> **[Results pending]** — loss curves, held-out perplexity, and a Polish eval (PolNative) to go here. The claim we are validating, à la the Dynaword paper's training experiments: clean, openly-licensed data is enough to train a capable Polish LM.
+
+## 5. Why it matters
+
+Put together, this is a Polish language-model stack where you can point at **every** component — the data, the tokenizer, the training run — and trace it to a license that permits the use. No scraped text of unknown rights. No copyrighted media. That is the spirit of the Open Source AI definition, applied end-to-end, for a language that didn't have it.
+
+## 6. Release
+
+- **Corpus:** `SlayerLab/polish-dynaword` (HF dataset, CC-BY-SA-4.0; per-source datasheets).
+- **Tokenizer:** `polish_bpe_32k.json`.
+- **Code:** reproducible build + tokenization + training scripts.
+
+*Built on the shoulders of the Dynaword guidelines (Enevoldsen et al., 2025) and the open-data communities behind every source.*

CHANGELOG.md

@@ -0,0 +1,7 @@
+# Changelog
+
+## v0.2.0 (2026-06-15)
+
+- Initial release: 11 openly-licensed sources, 2,490,773 docs, 6.22B tokens (tiktoken proxy).
+- Sources: eurlex, parliamentary, wikisource, wikipedia, dziennik_ustaw, wolne_lektury, wikiquote, eltec_pol, wikivoyage, wikibooks, wikinews.
+- Excluded (see README): open_subtitles_corpus, europeana_eu_pl_corpus, 1000_novels_corpus_CLARIN-PL, project_gutenberg_pl_corpus.

LICENSE

@@ -0,0 +1,5 @@
+Polish DynaWord is released under Creative Commons Attribution-ShareAlike 4.0
+International (CC-BY-SA-4.0): https://creativecommons.org/licenses/by-sa/4.0/
+
+Per-source upstream licenses and attribution are documented in each
+data/<source>/<source>.md datasheet.

PAPER.md

@@ -0,0 +1,240 @@
+# Polish DynaWord: An Openly-Licensed, Traceable, Continuously-Developed Pretraining Corpus for Polish
+
+**Kacper Wikieł**
+SlayerLab · `k.wikiel@gmail.com`
+
+---
+
+## Abstract
+
+We present **Polish DynaWord**, an openly-licensed, human-authored text corpus for
+Polish, built as a Polish edition of the *Dynaword* family (Enevoldsen et al.,
+2025, arXiv:2508.02271). The corpus follows four principles inherited from
+Dynaword — **open and traceable licensing, reproducibility, per-source
+documentation, and versioned extensibility** — and explicitly treats *curation*,
+not raw text collection, as its contribution. Version 0.2.0 comprises **2,490,773
+documents and ≈6.22 billion tokens** (tiktoken `cl100k` proxy) drawn from eleven
+openly-licensed Polish sources, each vetted for a *documented* legal basis rather
+than a vague "public domain" claim. Five sources are openly-licensed corpora
+redistributed via SpeakLeash; the remaining six are fetched *directly* from
+upstream — Polish primary legislation (Dziennik Ustaw / Monitor Polski) via the
+Sejm ELI API, four Wikimedia sibling projects (Wikinews, Wikivoyage, Wikibooks,
+Wikiquote) via the official dumps, and the ELTeC-pol literary collection via
+GitHub — each with full per-document provenance. Beyond scale, this broadens the
+*register* coverage from encyclopedic/legal/literary-classic text toward
+news, travel, instructional and quotation prose. We report the construction
+pipeline, the per-source legal review (including the sources we deliberately
+*exclude* and why), and the dataset statistics. The corpus exceeds the 4.8B-token
+Danish Dynaword in scale while maintaining a comparably conservative licensing
+posture.
+
+---
+
+## 1. Introduction
+
+Large-language-model pretraining for languages other than English is constrained
+less by modeling technique than by the availability of **legally clean** text.
+Many widely-used corpora are assembled from web crawls of uncertain provenance,
+mixing public-domain works, openly-licensed material, and in-copyright content
+without per-document accounting. The *Dynaword* project (Enevoldsen et al., 2025)
+proposed an alternative: a continuously-developed corpus where every source is
+*openly licensed* and that license is *traceable* to a documented legal basis,
+with the whole pipeline reproducible and each source individually documented as a
+datasheet (Gebru et al., 2021).
+
+Polish is comparatively well-served for raw text — the SpeakLeash project has
+aggregated 800+ Polish datasets — but most of that material carries "conditional"
+licensing (web-crawl or non-commercial terms) that is unsuitable for an
+openly-licensed release. **Polish DynaWord** addresses this gap. Its contribution
+is not the collection of new text but the *curation*: a per-source license
+review, a uniform minimal-filtering pipeline, transparent documentation of both
+included and excluded sources, and a reproducible build.
+
+## 2. Related work
+
+**Dynaword and the Danish lineage.** Polish DynaWord directly adapts the Danish
+Dynaword (Enevoldsen et al., 2025), itself a successor to the Danish Gigaword
+project. A key lesson carried over is the treatment of *derivative* content: the
+Danish effort excluded OpenSubtitles because subtitle text is a derivative work of
+copyrighted film/TV dialogue. We apply the same reasoning (§6).
+
+**SpeakLeash / Bielik.** SpeakLeash is the principal open aggregator of Polish
+text and the redistribution channel for most of our included sources. We use
+SpeakLeash as an *intermediate aggregator* and preserve upstream
+license/attribution; SpeakLeash is credited accordingly. Crucially, SpeakLeash's
+redistribution does not itself confer a license — the right to redistribute flows
+from each source's upstream license, which we document per source.
+
+**Datasheets for Datasets.** Following Gebru et al. (2021), each source ships a
+datasheet recording provenance, licensing basis, domain, time range, and the
+filters applied.
+
+## 3. Design principles
+
+1. **Open & traceable licensing.** Every source must be openly licensed *and*
+   carry a documented legal basis (statutory exemption or explicit upstream
+   license), not an unexplained "public domain" assertion.
+2. **Reproducibility.** A single command rebuilds the corpus from source archives
+   (`src/build_dynaword.py`).
+3. **Documented.** One datasheet per source; aggregate statistics in the dataset
+   card.
+4. **Extensibility.** Versioned; new sources and removals are tracked in a
+   changelog, and sources can be added via native fetchers (§8).
+
+## 4. Sources and licensing
+
+Version 0.2.0 includes eleven sources. For each we record the upstream origin and
+the *traceable basis* for its license.
+
+| Source | Domain | License | Traceable basis |
+|---|---|---|---|
+| Polish Wikipedia | encyclopedic | CC-BY-SA-3.0 | Wikimedia dumps released under CC-BY-SA 3.0; attribution + share-alike preserved |
+| Polish Wikisource | source texts | CC-BY-SA-3.0 | Wikimedia dumps; underlying works public-domain, community-transcribed |
+| EUR-Lex (PL) | EU legislation | CC-BY-4.0 | Commission Decision 2011/833/EU; normative acts outside copyright (PL art. 4 *pr. aut.*) |
+| Polish Parliamentary Corpus (Sejm/Senat) | political/spoken | public-domain / CC-BY-4.0 | Official documents outside copyright (PL art. 4 *pr. aut.*); redistributed by IPI PAN |
+| Wolne Lektury | literature | CC-BY-SA-4.0 / Wolna Sztuka 1.3 | Published under CC-BY-SA 4.0 or Free Art License; PD classics + cleared works |
+| Polish Wikinews | news | CC-BY-2.5 | Wikinews released under CC-BY 2.5; attribution preserved |
+| Polish Wikivoyage | travel | CC-BY-SA-3.0 | Wikivoyage released under CC-BY-SA 3.0 |
+| Polish Wikibooks | instructional | CC-BY-SA-3.0 | Wikibooks released under CC-BY-SA 3.0 |
+| Polish Wikiquote | quotations | CC-BY-SA-3.0 | Wikiquote released under CC-BY-SA 3.0 |
+| ELTeC-pol | literature (1840–1920) | CC-BY-4.0 | COST Action CA16204 release under CC-BY 4.0 |
+| Dziennik Ustaw + Monitor Polski | national legislation | public-domain (CC0) | Normative acts not subject to copyright (PL art. 4 *pr. aut.*); fetched directly from the Sejm ELI API |
+
+Because two included sources (Wikipedia/Wikisource, Wolne Lektury) are
+share-alike, the **aggregate corpus is released under CC-BY-SA-4.0**, with
+attribution due to each upstream source and to SpeakLeash as the aggregator.
+
+## 5. Construction methodology
+
+**Provenance.** Source archives are obtained from SpeakLeash's public
+redistribution bucket (`speakleash-ds-pub`) in `jsonl.zst` form. The build streams
+each archive, applies gates, and writes one Parquet file per source.
+
+**Filters (intentionally minimal).** Consistent with Dynaword guidelines, heavy
+quality filtering and mixture-weighting are deferred to downstream training; the
+build applies only:
+
+- drop documents shorter than 200 characters;
+- drop non-Polish text (diacritic-to-letter ratio threshold);
+- exact cross-source deduplication (SHA-1 over normalized text), with earlier
+  sources winning duplicates;
+- for OCR-bearing sources, an alpha-character-ratio gate against OCR garble.
+
+**Schema.** Every document is stored as
+`(id, text, source, added, created, token_count)`. Token counts are a fast
+tiktoken `cl100k` proxy (within ≈1% of a Llama-3 tokenizer); a canonical Llama-3
+recount is planned for a tagged release.
+
+## 6. Statistics
+
+| Source | Documents | Characters | Tokens (proxy) |
+|---|---:|---:|---:|
+| Polish Wikipedia | 1,171,897 | 1.85 B | 707.2 M |
+| Polish Wikisource | 632,005 | 1.97 B | 801.9 M |
+| EUR-Lex (PL) | 243,060 | 5.98 B | 2,378.1 M |
+| Polish Parliamentary Corpus | 324,622 | 4.49 B | 1,646.8 M |
+| Wolne Lektury | 6,141 | 0.26 B | 103.0 M |
+| Polish Wikinews | 24,386 | 0.03 B | 12.1 M |
+| Polish Wikivoyage | 13,645 | 0.05 B | 17.1 M |
+| Polish Wikibooks | 9,112 | 0.04 B | 15.6 M |
+| Polish Wikiquote | 30,363 | 0.08 B | 31.9 M |
+| ELTeC-pol | 100 | 0.05 B | 21.5 M |
+| Dziennik Ustaw + Monitor Polski | 35,442 | 1.23 B | 486.1 M |
+| **Total (v0.2.0)** | **2,490,773** | **16.0 B** | **6.22 B** |
+
+For comparison, the Danish Dynaword reports ≈4.8 B tokens; Polish DynaWord v0.2.0
+is larger while remaining within a strictly openly-licensed source set. The full
+build completes in ≈4 minutes on 14 CPU cores. The five sources added in v0.2.0
+contribute modest volume (≈98 M tokens) but broaden register coverage — news,
+travel, instructional, quotation and modern literary prose — addressing the
+encyclopedic/legal skew of the earlier versions.
+
+## 7. Excluded sources (curation as contribution)
+
+Transparency about what is *left out* is part of the editorial contribution. Each
+exclusion records a stated legal reason:
+
+- **OpenSubtitles (PL)** — subtitle text is a derivative work of copyrighted
+  film/TV dialogue and largely unlicensed; excluded on the Danish Gigaword
+  precedent.
+- **Europeana (PL)** — an aggregation of items with *mixed per-record rights*
+  (PD / CC-BY-NC / rights-reserved). The SpeakLeash redistribution does not carry
+  the `edm:rights` field, so per-record filtering is impossible without external
+  enrichment; excluded pending that work (≈203k docs / ≈5.2B tokens forgone).
+- **1000 Novels (CLARIN-PL)** — labelled CC-BY-4.0, but the collection likely
+  includes in-copyright contemporary works; excluded pending per-title verification.
+- **Project Gutenberg (PL)** — only 31 Polish books (4.3 MB), near-redundant with
+  Wikisource/Wolne Lektury, and claiming public domain *in the US only*; dropped
+  to avoid a per-work PD-in-EU check for negligible token gain.
+
+## 8. Native sources fetched directly from upstream
+
+Beyond the SpeakLeash-redistributed corpora, six sources are fetched *directly*
+from their publishers. Direct sourcing strengthens provenance (each document
+records its upstream URL/identifier), keeps the data fresh, and distinguishes the
+corpus from a mirror of any single aggregator. Each is implemented as a small,
+reproducible fetcher under `src/`.
+
+**Polish primary legislation** (`src/fetch_eli.py`). We fetch **Dziennik Ustaw**
+and **Monitor Polski** from the Sejm ELI API (`api.sejm.gov.pl/eli/acts`),
+retrieving each act's HTML and converting it to plain text. For v0.2.0 we include
+acts marked in force over 1990–2025: **35,442 documents (≈486 M tokens)**.
+Normative acts and their official drafts are **not subject to copyright** under
+Polish law (art. 4 of the Act on Copyright and Related Rights). This adds Polish
+*national* law, complementing the EU law present via EUR-Lex.
+
+**Wikimedia sibling projects** (`src/fetch_wikimedia.py`). We fetch the Polish
+**Wikinews**, **Wikivoyage**, **Wikibooks** and **Wikiquote** from the official
+Wikimedia dumps, keeping mainspace pages and stripping wikitext to plain text with
+`mwparserfromhell`. These contribute the **news, travel, instructional and
+quotation** registers (≈77 k documents, ≈77 M tokens combined) under CC-BY /
+CC-BY-SA, compatible with the aggregate license.
+
+**ELTeC-pol** (`src/fetch_eltec.py`). We fetch the Polish part of the European
+Literary Text Collection (COST Action CA16204) from GitHub — **100 novels**
+(1840–1920, ≈21.5 M tokens) released CC-BY-4.0 — adding modern narrative literary
+prose distinct from the older texts in Wikisource and Wolne Lektury.
+
+## 9. Ethics, licensing and limitations
+
+**Personal data.** The corpus contains only text its sources already published
+openly or as official record; it therefore includes names and statements of
+public figures acting in a public capacity (parliamentary speakers, authorities
+named in legislation, people described in encyclopedic articles). No private,
+non-public personal data is collected. A notice-and-takedown procedure allows
+data subjects and rightsholders to request removal from subsequent versions.
+
+**License compliance.** The release is CC-BY-SA-4.0; downstream users must satisfy
+attribution and share-alike obligations for derivatives, and attribute the
+upstream sources and SpeakLeash. Per-source licenses are reproduced in good faith
+as documented by upstream and by SpeakLeash; we make no independent legal warranty
+about any individual document's copyright status.
+
+**Limitations.** (i) Token counts are a tiktoken proxy pending a canonical Llama-3
+recount. (ii) Evaluation-set decontamination (n-gram overlap against Polish
+benchmarks) is *not yet applied* and is required before the corpus is used to
+train models reported on those benchmarks. (iii) Deduplication is exact-match
+only; near-duplicate removal is left to downstream use. (iv) Gutenberg-style
+PD-in-EU edge cases are handled by exclusion rather than per-work adjudication.
+
+## 10. Availability
+
+The dataset, per-source datasheets, build/fetch code, and documentation are
+released at `huggingface.co/datasets/SlayerLab/polish-dynaword` (CC-BY-SA-4.0).
+The corpus is versioned; this paper describes v0.2.0.
+
+## References
+
+- Enevoldsen, K. et al. (2025). *Dynaword: A Continuously Developed, Openly
+  Licensed Corpus.* arXiv:2508.02271.
+- Gebru, T. et al. (2021). *Datasheets for Datasets.* Communications of the ACM.
+- SpeakLeash / Bielik project. *Open Polish text datasets.* speakleash.org.
+- Strømberg-Derczynski, L. et al. (2021). *The Danish Gigaword Corpus.* NoDaLiDa.
+- Sejm RP. *ELI API — Internetowy System Aktów Prawnych.* api.sejm.gov.pl/eli.
+- Ustawa z dnia 4 lutego 1994 r. o prawie autorskim i prawach pokrewnych, art. 4.
+
+---
+
+*Draft — generated alongside the v0.2.0 corpus release. Numbers reflect the
+eleven-source build (five SpeakLeash-redistributed + six fetched directly from
+upstream). Evaluation-set decontamination (§9) is still pending.*

README.md

@@ -0,0 +1,126 @@
+---
+license: cc-by-sa-4.0
+language:
+- pl
+pretty_name: Polish DynaWord
+task_categories:
+- text-generation
+size_categories:
+- 1M<n<10M
+tags:
+- polish
+- pretraining
+- dynaword
+---
+
+# Polish DynaWord
+
+A continuously developed, **openly-licensed**, human-text Polish corpus — a Polish
+edition in the [Dynaword](https://huggingface.co/datasets/danish-foundation-models/danish-dynaword)
+family (Enevoldsen et al., [arXiv:2508.02271](https://arxiv.org/abs/2508.02271)).
+
+> **v0.2.0** · 2,490,773 documents · **6.22B tokens** (tiktoken proxy;
+> canonical Llama-3 count at release) · 11 sources
+
+## What this dataset contributes
+The raw texts come from existing open corpora (redistributed via SpeakLeash and,
+where applicable, fetched from upstream). **The value added here is the curation,
+not the bytes**, following the Dynaword methodology:
+
+1. **License review per source** — each source vetted for an *openly-licensed,
+   traceable* legal basis (documented in its datasheet); sources that fail the
+   review are **excluded with a stated reason** (see table below), not silently
+   kept. This is the core editorial work.
+2. **Filtering & normalization** — minimal, reproducible gates (short-doc,
+   non-Polish, exact cross-source dedup, OCR garble) applied uniformly to one
+   clean schema: `id, text, source, added, created, token_count`.
+3. **Documentation** — a datasheet per source (Gebru et al. 2021) + this card,
+   so provenance and licensing are auditable rather than assumed.
+4. **Reproducibility & versioning** — `src/` rebuilds the corpus from sources;
+   new sources and removals are tracked in the CHANGELOG.
+
+Credit for the underlying texts belongs to the upstream sources and to SpeakLeash
+as the redistributing aggregator; this release does not claim ownership of them
+(see Disclaimer).
+
+## Guiding principles
+1. **Open & traceable licensing** — every source is *openly licensed* with a documented
+   legal basis (see each datasheet's "traceable basis"), not a vague "public domain".
+2. **Reproducibility** — `src/build_dynaword.py` rebuilds the corpus from sources.
+3. **Documented** — a datasheet per source under `data/<source>/`.
+4. **Extensibility** — versioned; new sources via PR.
+
+## Sources
+| source | description | license | documents | tokens |
+|---|---|---|---:|---:|
+| [eurlex](data/eurlex/eurlex.md) | EUR-Lex (EU legal acts, Polish) | `CC-BY-4.0` | 243,060 | 2,378.1M |
+| [parliamentary](data/parliamentary/parliamentary.md) | Polish Parliamentary Corpus (Sejm/Senat) | `public-domain (official documents)` | 324,622 | 1,646.8M |
+| [wikisource](data/wikisource/wikisource.md) | Polish Wikisource | `CC-BY-SA-3.0` | 632,005 | 801.9M |
+| [wikipedia](data/wikipedia/wikipedia.md) | Polish Wikipedia | `CC-BY-SA-3.0` | 1,171,897 | 707.2M |
+| [dziennik_ustaw](data/dziennik_ustaw/dziennik_ustaw.md) | Dziennik Ustaw + Monitor Polski (Polish primary legislation) | `public-domain (official documents)` | 35,442 | 486.1M |
+| [wolne_lektury](data/wolne_lektury/wolne_lektury.md) | Wolne Lektury (school readings) | `CC-BY-SA-4.0 / Wolna Sztuka 1.3` | 6,141 | 103.0M |
+| [wikiquote](data/wikiquote/wikiquote.md) | Polish Wikiquote (quotations) | `CC-BY-SA-3.0` | 30,363 | 31.9M |
+| [eltec_pol](data/eltec_pol/eltec_pol.md) | ELTeC-pol (European Literary Text Collection, Polish) | `CC-BY-4.0` | 100 | 21.5M |
+| [wikivoyage](data/wikivoyage/wikivoyage.md) | Polish Wikivoyage (travel guides) | `CC-BY-SA-3.0` | 13,645 | 17.1M |
+| [wikibooks](data/wikibooks/wikibooks.md) | Polish Wikibooks (open textbooks) | `CC-BY-SA-3.0` | 9,112 | 15.6M |
+| [wikinews](data/wikinews/wikinews.md) | Polish Wikinews | `CC-BY-2.5` | 24,386 | 12.1M |
+| **total** | | | **2,490,773** | **6,221.4M** |
+
+## Method
+Only **human-authored** text — no synthetic, machine-translated, or auto-transcribed
+data. Gates are intentionally minimal (drop short docs, non-Polish, exact duplicates,
+OCR garble); heavy quality filtering and mix-weighting are left to downstream training.
+Evaluation-set decontamination is applied/marked separately. Schema:
+`id, text, source, added, created, token_count`.
+
+## Excluded sources (transparency)
+Sources we reviewed and **deliberately left out** — part of the curation:
+
+| source | reason |
+|---|---|
+| `open_subtitles_corpus` | Derivative of copyrighted film/TV dialogue; OpenSubtitles uploads largely unlicensed. Same copyright lesson as Danish Gigaword's OpenSubtitles (paper 2508.02271). Not openly licensed. |
+| `europeana_eu_pl_corpus` | Aggregated items with mixed per-record rights (PD / CC-BY-NC / rights-reserved). Needs per-record edm:rights filter before any inclusion. |
+| `1000_novels_corpus_CLARIN-PL` | CC-BY-4.0 label, but 'novels' likely include in-copyright contemporary works; verify titles/years on CLARIN handle 11321/312 before inclusion. |
+| `project_gutenberg_pl_corpus` | Only 31 PL books (4.3MB) — PG is ~99% English; Polish PD literature already covered by wolne_lektury + wikisource (so near-redundant after dedup). Dropped to avoid the PD-in-EU per-work check (PG claims PD-in-US only) for negligible token gain. |
+
+## Personal & sensitive data
+This corpus contains **only** text that its upstream sources already published
+under open licenses or as official public-domain record. It therefore includes
+names and statements of **public figures acting in a public capacity** — e.g.
+parliamentary speakers (PPC), authorities named in legal acts (EUR-Lex), and
+people described in encyclopedic articles (Wikipedia/Wikisource). No private,
+non-public personal data was collected or added. If you are a data subject and
+want content concerning you removed, contact **k.wikiel@gmail.com** — it will be dropped
+from the next version (see retroactive-removal policy below).
+
+## Disclaimer & legal
+- **Provenance in good faith.** Per-source licenses are reproduced *as documented
+  by the upstream sources and by SpeakLeash* (the intermediate aggregator), to the
+  best of our knowledge. We make no independent legal warranty about the copyright
+  status of any individual document.
+- **No ownership claim.** This release is a *curated, license-reviewed, documented
+  aggregation*. We claim no ownership of the underlying texts; rights remain with
+  the original authors/rightsholders under their respective licenses.
+- **Provided "as is"**, without warranty of any kind, express or implied. This is
+  not legal advice.
+- **Your compliance is yours.** Downstream users must satisfy each upstream
+  license themselves — in particular **CC-BY-SA-4.0 attribution and share-alike**
+  for derivatives of this dataset, and attribution to the upstream sources and to
+  SpeakLeash.
+- **Notice-and-takedown.** Any source or rightsholder raising a substantiated
+  objection can have material removed: contact **k.wikiel@gmail.com**; it is dropped from
+  the next version and recorded in the CHANGELOG. Removal is retroactive
+  going-forward (prior immutable snapshots/commits may persist).
+
+## License & attribution
+Released under **CC-BY-SA-4.0** (copyleft inherited from CC-BY-SA sources such as
+Wikipedia/Wikisource/Wolne Lektury). Attribution due to each upstream (see datasheets)
+and to **SpeakLeash** as the intermediate aggregator. Retroactive-removal policy: a
+source that raises an objection is dropped from subsequent versions, recorded in the
+CHANGELOG.
+
+## Reproduce
+```bash
+python3 src/build_dynaword.py --all --speakleash-dir <speakleash_zst_dir> --out .
+python3 src/make_docs.py
+```

data/dziennik_ustaw/dziennik_ustaw.md

@@ -0,0 +1,33 @@
+# dziennik_ustaw
+
+Dziennik Ustaw + Monitor Polski (Polish primary legislation)
+
+## Dataset description
+- **Source (upstream):** https://api.sejm.gov.pl/eli/acts
+- **Domain:** legal
+- **Language:** Polish (pl)
+- **License:** `public-domain (official documents)`
+- **Created (range):** 1918-01-01, 2025-12-31
+- **Added:** 2026-06-15
+
+## Licensing — traceable basis
+Normative acts and their official drafts are not subject to copyright (PL art. 4 ustawy o prawie autorskim). Published by the state; reuse unrestricted.
+
+## Provenance
+Fetched directly from the Sejm ELI API (text.html per act) by src/fetch_eli.py — a native source, not redistributed via SpeakLeash. Full provenance: each act's `meta.url` is its ELI address, resolvable on api.sejm.gov.pl.
+
+## Statistics
+| documents | characters | tokens (tiktoken proxy) |
+|---:|---:|---:|
+| 35,442 | 1,226,137,838 | 486,126,575 |
+
+## Filters applied (build_dynaword.py)
+Minimal, per Dynaword guidelines (heavy filtering left to downstream use):
+- drop documents < 200 chars: **0**
+- drop non-Polish (diacritic ratio): **0**
+- exact cross-source dedup (sha1): **0**
+- OCR alpha-ratio < 0.70 (OCR sources only): **0**
+- read 35,442 → kept 35,442
+
+Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
+Llama-3 count is computed at release.

data/dziennik_ustaw/dziennik_ustaw.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3cf60c3d3298390bd46e4feadd7b581d79141a1b99ff0ea06645012bbc096b65
+size 313228363

data/dziennik_ustaw/dziennik_ustaw.stats.json

@@ -0,0 +1,12 @@
+{
+  "read": 35442,
+  "kept": 35442,
+  "drop_short": 0,
+  "drop_lang": 0,
+  "drop_dup": 0,
+  "drop_ocr": 0,
+  "chars": 1226137838,
+  "tokens": 486126575,
+  "secs": 17.8,
+  "license": "public-domain (official documents)"
+}

data/eltec_pol/eltec_pol.md

@@ -0,0 +1,33 @@
+# eltec_pol
+
+ELTeC-pol (European Literary Text Collection, Polish)
+
+## Dataset description
+- **Source (upstream):** https://github.com/COST-ELTeC/ELTeC-pol
+- **Domain:** literature
+- **Language:** Polish (pl)
+- **License:** `CC-BY-4.0`
+- **Created (range):** 1840-01-01, 1920-12-31
+- **Added:** 2026-06-15
+
+## Licensing — traceable basis
+ELTeC corpora are released under CC-BY 4.0 by the COST Action 'Distant Reading for European Literary History' (CA16204).
+
+## Provenance
+Fetched directly from the COST-ELTeC/ELTeC-pol GitHub repository (level1 TEI) by src/fetch_eltec.py; reading text extracted from TEI <body> paragraphs.
+
+## Statistics
+| documents | characters | tokens (tiktoken proxy) |
+|---:|---:|---:|
+| 100 | 54,369,249 | 21,486,720 |
+
+## Filters applied (build_dynaword.py)
+Minimal, per Dynaword guidelines (heavy filtering left to downstream use):
+- drop documents < 200 chars: **0**
+- drop non-Polish (diacritic ratio): **0**
+- exact cross-source dedup (sha1): **0**
+- OCR alpha-ratio < 0.70 (OCR sources only): **0**
+- read 100 → kept 100
+
+Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
+Llama-3 count is computed at release.

data/eurlex/eurlex.md

@@ -0,0 +1,33 @@
+# eurlex
+
+EUR-Lex (EU legal acts, Polish)
+
+## Dataset description
+- **Source (upstream):** https://eur-lex.europa.eu
+- **Domain:** legal
+- **Language:** Polish (pl)
+- **License:** `CC-BY-4.0`
+- **Created (range):** 1952-01-01, 2024-12-31
+- **Added:** 2026-06-15
+
+## Licensing — traceable basis
+EU documents reusable under Commission Decision 2011/833/EU; normative acts fall outside copyright (PL art. 4 pr. aut.). Reuse authorised with source acknowledgement.
+
+## Provenance
+Pulled from SpeakLeash's public redistribution (`speakleash-ds-pub`, key `eurlex_corpus`) of the upstream source above. SpeakLeash credited as intermediate aggregator; upstream license/attribution preserved.
+
+## Statistics
+| documents | characters | tokens (tiktoken proxy) |
+|---:|---:|---:|
+| 243,060 | 5,976,949,249 | 2,378,055,718 |
+
+## Filters applied (build_dynaword.py)
+Minimal, per Dynaword guidelines (heavy filtering left to downstream use):
+- drop documents < 200 chars: **0**
+- drop non-Polish (diacritic ratio): **1,390**
+- exact cross-source dedup (sha1): **1,825**
+- OCR alpha-ratio < 0.70 (OCR sources only): **0**
+- read 246,275 → kept 243,060
+
+Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
+Llama-3 count is computed at release.

data/eurlex/eurlex.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9da0ff96941150af85960497e1aab0188f7956d485e41f7170fa402346961fb1
+size 1692243551

data/eurlex/eurlex.stats.json

@@ -0,0 +1,12 @@
+{
+  "read": 246275,
+  "kept": 243060,
+  "drop_short": 0,
+  "drop_lang": 1390,
+  "drop_dup": 1825,
+  "drop_ocr": 0,
+  "chars": 5976949249,
+  "tokens": 2378055718,
+  "secs": 69.8,
+  "license": "CC-BY-4.0"
+}

data/parliamentary/parliamentary.md

@@ -0,0 +1,33 @@
+# parliamentary
+
+Polish Parliamentary Corpus (Sejm/Senat)
+
+## Dataset description
+- **Source (upstream):** http://clip.ipipan.waw.pl/PPC
+- **Domain:** political/spoken
+- **Language:** Polish (pl)
+- **License:** `public-domain (official documents)`
+- **Created (range):** 1991-01-01, 2019-12-31
+- **Added:** 2026-06-15
+
+## Licensing — traceable basis
+Parliamentary proceedings are official documents outside copyright (PL art. 4 pr. aut.); PPC redistributed by IPI PAN under CC-BY-4.0.
+
+## Provenance
+Pulled from SpeakLeash's public redistribution (`speakleash-ds-pub`, key `PPC_corpus`) of the upstream source above. SpeakLeash credited as intermediate aggregator; upstream license/attribution preserved.
+
+## Statistics
+| documents | characters | tokens (tiktoken proxy) |
+|---:|---:|---:|
+| 324,622 | 4,492,746,014 | 1,646,835,986 |
+
+## Filters applied (build_dynaword.py)
+Minimal, per Dynaword guidelines (heavy filtering left to downstream use):
+- drop documents < 200 chars: **583**
+- drop non-Polish (diacritic ratio): **0**
+- exact cross-source dedup (sha1): **18,592**
+- OCR alpha-ratio < 0.70 (OCR sources only): **0**
+- read 343,797 → kept 324,622
+
+Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
+Llama-3 count is computed at release.

data/parliamentary/parliamentary.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c5cc1c90ba06f250dbb272ae472d1cb7da9fe101ca8fdd3ea3dcba14c4eebef
+size 1732860869

data/parliamentary/parliamentary.stats.json

@@ -0,0 +1,12 @@
+{
+  "read": 343797,
+  "kept": 324622,
+  "drop_short": 583,
+  "drop_lang": 0,
+  "drop_dup": 18592,
+  "drop_ocr": 0,
+  "chars": 4492746014,
+  "tokens": 1646835986,
+  "secs": 48.8,
+  "license": "public-domain (official documents)"
+}

data/wikibooks/wikibooks.md

@@ -0,0 +1,33 @@
+# wikibooks
+
+Polish Wikibooks (open textbooks)
+
+## Dataset description
+- **Source (upstream):** https://pl.wikibooks.org/
+- **Domain:** instructional
+- **Language:** Polish (pl)
+- **License:** `CC-BY-SA-3.0`
+- **Created (range):** 2004-01-01, 2025-12-31
+- **Added:** 2026-06-15
+
+## Licensing — traceable basis
+Wikibooks content is released under CC-BY-SA 3.0; attribution + share-alike preserved.
+
+## Provenance
+Fetched directly from the official Wikimedia dump (plwikibooks-latest-pages-articles) by src/fetch_wikimedia.py; mainspace only, wikitext stripped via mwparserfromhell.
+
+## Statistics
+| documents | characters | tokens (tiktoken proxy) |
+|---:|---:|---:|
+| 9,112 | 38,892,589 | 15,571,295 |
+
+## Filters applied (build_dynaword.py)
+Minimal, per Dynaword guidelines (heavy filtering left to downstream use):
+- drop documents < 200 chars: **0**
+- drop non-Polish (diacritic ratio): **539**
+- exact cross-source dedup (sha1): **15**
+- OCR alpha-ratio < 0.70 (OCR sources only): **0**
+- read 9,666 → kept 9,112
+
+Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
+Llama-3 count is computed at release.

data/wikinews/wikinews.md

@@ -0,0 +1,33 @@
+# wikinews
+
+Polish Wikinews
+
+## Dataset description
+- **Source (upstream):** https://pl.wikinews.org/
+- **Domain:** news
+- **Language:** Polish (pl)
+- **License:** `CC-BY-2.5`
+- **Created (range):** 2005-01-01, 2025-12-31
+- **Added:** 2026-06-15
+
+## Licensing — traceable basis
+Wikinews content is released under CC-BY 2.5; attribution to contributors preserved.
+
+## Provenance
+Fetched directly from the official Wikimedia dump (plwikinews-latest-pages-articles) by src/fetch_wikimedia.py; mainspace only, wikitext stripped via mwparserfromhell.
+
+## Statistics
+| documents | characters | tokens (tiktoken proxy) |
+|---:|---:|---:|
+| 24,386 | 32,524,646 | 12,141,355 |
+
+## Filters applied (build_dynaword.py)
+Minimal, per Dynaword guidelines (heavy filtering left to downstream use):
+- drop documents < 200 chars: **0**
+- drop non-Polish (diacritic ratio): **6**
+- exact cross-source dedup (sha1): **0**
+- OCR alpha-ratio < 0.70 (OCR sources only): **0**
+- read 24,392 → kept 24,386
+
+Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
+Llama-3 count is computed at release.

data/wikipedia/wikipedia.md

@@ -0,0 +1,33 @@
+# wikipedia
+
+Polish Wikipedia
+
+## Dataset description
+- **Source (upstream):** https://dumps.wikimedia.org/plwiki/latest/
+- **Domain:** encyclopedic
+- **Language:** Polish (pl)
+- **License:** `CC-BY-SA-3.0`
+- **Created (range):** 2001-01-15, 2026-06-14
+- **Added:** 2026-06-15
+
+## Licensing — traceable basis
+Wikimedia dumps explicitly released under CC-BY-SA 3.0; attribution to contributors + share-alike preserved.
+
+## Provenance
+Pulled from SpeakLeash's public redistribution (`speakleash-ds-pub`, key `plwiki`) of the upstream source above. SpeakLeash credited as intermediate aggregator; upstream license/attribution preserved.
+
+## Statistics
+| documents | characters | tokens (tiktoken proxy) |
+|---:|---:|---:|
+| 1,171,897 | 1,844,707,761 | 707,194,207 |
+
+## Filters applied (build_dynaword.py)
+Minimal, per Dynaword guidelines (heavy filtering left to downstream use):
+- drop documents < 200 chars: **296,738**
+- drop non-Polish (diacritic ratio): **1,172**
+- exact cross-source dedup (sha1): **113**
+- OCR alpha-ratio < 0.70 (OCR sources only): **0**
+- read 1,469,920 → kept 1,171,897
+
+Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
+Llama-3 count is computed at release.

data/wikipedia/wikipedia.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f66fac8b3c4da41ff8ccbe7c568e38ca710ae42f85d409a9702503eb99dd8185
+size 866297438

data/wikipedia/wikipedia.stats.json

@@ -0,0 +1,12 @@
+{
+  "read": 1469920,
+  "kept": 1171897,
+  "drop_short": 296738,
+  "drop_lang": 1172,
+  "drop_dup": 113,
+  "drop_ocr": 0,
+  "chars": 1844707761,
+  "tokens": 707194207,
+  "secs": 20.6,
+  "license": "CC-BY-SA-3.0"
+}

data/wikiquote/wikiquote.md

@@ -0,0 +1,33 @@
+# wikiquote
+
+Polish Wikiquote (quotations)
+
+## Dataset description
+- **Source (upstream):** https://pl.wikiquote.org/
+- **Domain:** quotations
+- **Language:** Polish (pl)
+- **License:** `CC-BY-SA-3.0`
+- **Created (range):** 2005-01-01, 2025-12-31
+- **Added:** 2026-06-15
+
+## Licensing — traceable basis
+Wikiquote content is released under CC-BY-SA 3.0; attribution + share-alike preserved.
+
+## Provenance
+Fetched directly from the official Wikimedia dump (plwikiquote-latest-pages-articles) by src/fetch_wikimedia.py; mainspace only, wikitext stripped via mwparserfromhell.
+
+## Statistics
+| documents | characters | tokens (tiktoken proxy) |
+|---:|---:|---:|
+| 30,363 | 82,338,997 | 31,896,591 |
+
+## Filters applied (build_dynaword.py)
+Minimal, per Dynaword guidelines (heavy filtering left to downstream use):
+- drop documents < 200 chars: **0**
+- drop non-Polish (diacritic ratio): **11**
+- exact cross-source dedup (sha1): **2**
+- OCR alpha-ratio < 0.70 (OCR sources only): **0**
+- read 30,376 → kept 30,363
+
+Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
+Llama-3 count is computed at release.

data/wikisource/wikisource.md

@@ -0,0 +1,33 @@
+# wikisource
+
+Polish Wikisource
+
+## Dataset description
+- **Source (upstream):** https://dumps.wikimedia.org/plwikisource/latest/
+- **Domain:** books/source-texts
+- **Language:** Polish (pl)
+- **License:** `CC-BY-SA-3.0`
+- **Created (range):** 2005-01-01, 2026-06-14
+- **Added:** 2026-06-15
+
+## Licensing — traceable basis
+Wikimedia dumps under CC-BY-SA 3.0; underlying texts are public-domain works transcribed by the community.
+
+## Provenance
+Pulled from SpeakLeash's public redistribution (`speakleash-ds-pub`, key `plwikisource`) of the upstream source above. SpeakLeash credited as intermediate aggregator; upstream license/attribution preserved.
+
+## Statistics
+| documents | characters | tokens (tiktoken proxy) |
+|---:|---:|---:|
+| 632,005 | 1,974,249,746 | 801,947,427 |
+
+## Filters applied (build_dynaword.py)
+Minimal, per Dynaword guidelines (heavy filtering left to downstream use):
+- drop documents < 200 chars: **0**
+- drop non-Polish (diacritic ratio): **15**
+- exact cross-source dedup (sha1): **0**
+- OCR alpha-ratio < 0.70 (OCR sources only): **0**
+- read 632,020 → kept 632,005
+
+Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
+Llama-3 count is computed at release.

data/wikisource/wikisource.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e066d5e43c84f0d70376e4162ff97f544b4e6122906458727277aa562efbff0
+size 984212691

data/wikisource/wikisource.stats.json

@@ -0,0 +1,12 @@
+{
+  "read": 632020,
+  "kept": 632005,
+  "drop_short": 0,
+  "drop_lang": 15,
+  "drop_dup": 0,
+  "drop_ocr": 0,
+  "chars": 1974249746,
+  "tokens": 801947427,
+  "secs": 62.0,
+  "license": "CC-BY-SA-3.0"
+}

data/wikivoyage/wikivoyage.md

@@ -0,0 +1,33 @@
+# wikivoyage
+
+Polish Wikivoyage (travel guides)
+
+## Dataset description
+- **Source (upstream):** https://pl.wikivoyage.org/
+- **Domain:** travel
+- **Language:** Polish (pl)
+- **License:** `CC-BY-SA-3.0`
+- **Created (range):** 2012-01-01, 2025-12-31
+- **Added:** 2026-06-15
+
+## Licensing — traceable basis
+Wikivoyage content is released under CC-BY-SA 3.0; attribution + share-alike preserved.
+
+## Provenance
+Fetched directly from the official Wikimedia dump (plwikivoyage-latest-pages-articles) by src/fetch_wikimedia.py; mainspace only, wikitext stripped via mwparserfromhell.
+
+## Statistics
+| documents | characters | tokens (tiktoken proxy) |
+|---:|---:|---:|
+| 13,645 | 45,184,300 | 17,128,200 |
+
+## Filters applied (build_dynaword.py)
+Minimal, per Dynaword guidelines (heavy filtering left to downstream use):
+- drop documents < 200 chars: **0**
+- drop non-Polish (diacritic ratio): **9**
+- exact cross-source dedup (sha1): **1**
+- OCR alpha-ratio < 0.70 (OCR sources only): **0**
+- read 13,655 → kept 13,645
+
+Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
+Llama-3 count is computed at release.

data/wolne_lektury/wolne_lektury.md

@@ -0,0 +1,33 @@
+# wolne_lektury
+
+Wolne Lektury (school readings)
+
+## Dataset description
+- **Source (upstream):** https://wolnelektury.pl/api/books/
+- **Domain:** literature
+- **Language:** Polish (pl)
+- **License:** `CC-BY-SA-4.0 / Wolna Sztuka 1.3`
+- **Created (range):** 1500-01-01, 2024-12-31
+- **Added:** 2026-06-15
+
+## Licensing — traceable basis
+Wolne Lektury publishes under CC-BY-SA 4.0 or Free Art License; texts are public-domain classics + cleared works.
+
+## Provenance
+Pulled from SpeakLeash's public redistribution (`speakleash-ds-pub`, key `wolne_lektury_corpus`) of the upstream source above. SpeakLeash credited as intermediate aggregator; upstream license/attribution preserved.
+
+## Statistics
+| documents | characters | tokens (tiktoken proxy) |
+|---:|---:|---:|
+| 6,141 | 261,892,000 | 103,009,797 |
+
+## Filters applied (build_dynaword.py)
+Minimal, per Dynaword guidelines (heavy filtering left to downstream use):
+- drop documents < 200 chars: **471**
+- drop non-Polish (diacritic ratio): **2**
+- exact cross-source dedup (sha1): **5**
+- OCR alpha-ratio < 0.70 (OCR sources only): **0**
+- read 6,619 → kept 6,141
+
+Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
+Llama-3 count is computed at release.

data/wolne_lektury/wolne_lektury.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:157556cc790b7909afbc544803af9fa18f8b4ed8b334f03cc65b36867c51e79d
+size 128537253

data/wolne_lektury/wolne_lektury.stats.json

@@ -0,0 +1,12 @@
+{
+  "read": 6619,
+  "kept": 6141,
+  "drop_short": 471,
+  "drop_lang": 2,
+  "drop_dup": 5,
+  "drop_ocr": 0,
+  "chars": 261892000,
+  "tokens": 103009797,
+  "secs": 10.8,
+  "license": "CC-BY-SA-4.0 / Wolna Sztuka 1.3"
+}

src/build_dynaword.py

@@ -0,0 +1,167 @@
+#!/usr/bin/env python3
+"""Build Polish DynaWord parquet shards from SpeakLeash .jsonl.zst sources.
+
+Parallel pipeline (uses all cores). Per source (paper 2508.02271, minimal gates):
+  stream jsonl.zst -> [workers: parse + Polish-lang check + drop-short +
+  OCR alpha-ratio + tiktoken token_count + sha1] -> [main: cross-source exact
+  dedup + id + parquet write].
+
+Sources processed in priority order so earlier sources win duplicates
+(wikipedia > wikisource > ...). Heavy quality filtering + mix-weighting are
+downstream (CPT), not here. token_count is a fast tiktoken proxy (~1% off
+Llama-3); canonical Llama-3 recount happens at release.
+
+Usage:
+  python3 src/build_dynaword.py --all --speakleash-dir ~/speakleash --out ~/dynaword
+  python3 src/build_dynaword.py --sources gutenberg --jobs 16
+"""
+from __future__ import annotations
+import argparse, hashlib, json, os, re, subprocess, sys, time
+from itertools import islice
+import multiprocessing as mp
+from pathlib import Path
+import pyarrow as pa
+import pyarrow.parquet as pq
+
+sys.path.insert(0, str(Path(__file__).resolve().parent))
+from sources import SOURCES, ADDED
+
+POLISH_RE = re.compile(r"[ąćęłńóśźżĄĆĘŁŃÓŚŹŻ]")
+ALPHA_RE = re.compile(r"[^\W\d_]", re.UNICODE)
+MIN_CHARS = 200
+MIN_POLISH_RATIO = 0.005
+MIN_ALPHA_RATIO = 0.70
+
+SCHEMA = pa.schema([
+    ("id", pa.string()), ("text", pa.string()), ("source", pa.string()),
+    ("added", pa.string()), ("created", pa.string()), ("token_count", pa.int64()),
+])
+
+_ENC = None  # per-worker tiktoken encoder
+
+
+def _init_worker():
+    global _ENC
+    import tiktoken
+    _ENC = tiktoken.get_encoding("cl100k_base")
+
+
+def _polish_ratio(text):
+    letters = ALPHA_RE.findall(text)
+    return len(POLISH_RE.findall(text)) / len(letters) if letters else 0.0
+
+
+def _process_chunk(args):
+    """Worker: gate + tokenize a batch of raw lines. Returns (records, stats)."""
+    is_ocr, created, lines = args
+    kept, texts = [], []
+    st = [0, 0, 0, 0]  # read, short, lang, ocr
+    for line in lines:
+        st[0] += 1
+        try:
+            text = (json.loads(line).get("text") or "").strip()
+        except Exception:
+            continue
+        if len(text) < MIN_CHARS:
+            st[1] += 1; continue
+        if _polish_ratio(text) < MIN_POLISH_RATIO:
+            st[2] += 1; continue
+        if is_ocr:
+            ar = len(ALPHA_RE.findall(text)) / len(text) if text else 0.0
+            if ar < MIN_ALPHA_RATIO:
+                st[3] += 1; continue
+        texts.append(text)
+    toks = [len(t) for t in _ENC.encode_ordinary_batch(texts, num_threads=1)] if texts else []
+    for t, tk in zip(texts, toks):
+        kept.append((t, created, tk, hashlib.sha1(t.encode("utf-8")).digest()))
+    return kept, st
+
+
+def _chunks(iterable, n):
+    it = iter(iterable)
+    while batch := list(islice(it, n)):
+        yield batch
+
+
+def build_source(name, cfg, sl_dir, out_root, pool, seen, counter):
+    src_path = sl_dir / f"{cfg.get('file_key', cfg.get('speakleash_key'))}.jsonl.zst"
+    if not src_path.exists():
+        print(f"  ! missing {src_path}"); return None
+    out_dir = out_root / "data" / name
+    out_dir.mkdir(parents=True, exist_ok=True)
+    writer = pq.ParquetWriter(out_dir / f"{name}.parquet", SCHEMA, compression="zstd")
+    st = {"read": 0, "kept": 0, "drop_short": 0, "drop_lang": 0,
+          "drop_dup": 0, "drop_ocr": 0, "chars": 0, "tokens": 0}
+    t0 = time.time()
+    is_ocr, created = bool(cfg.get("is_ocr")), cfg.get("created", "")
+    bid, btext, bcre, btok = [], [], [], []
+
+    def flush():
+        if not btext:
+            return
+        n = len(btext)
+        writer.write(pa.record_batch([
+            pa.array(bid), pa.array(btext), pa.array([name] * n),
+            pa.array([ADDED] * n), pa.array(bcre), pa.array(btok, pa.int64()),
+        ], schema=SCHEMA))
+        bid.clear(); btext.clear(); bcre.clear(); btok.clear()
+
+    proc = subprocess.Popen(["zstd", "-dc", str(src_path)],
+                            stdout=subprocess.PIPE, bufsize=1 << 22)
+    line_iter = (ln for ln in proc.stdout if ln.strip())
+    arg_iter = ((is_ocr, created, ch) for ch in _chunks(line_iter, 2000))
+    for kept, cst in pool.imap_unordered(_process_chunk, arg_iter, chunksize=1):
+        st["read"] += cst[0]; st["drop_short"] += cst[1]
+        st["drop_lang"] += cst[2]; st["drop_ocr"] += cst[3]
+        for text, cre, tok, h in kept:
+            if h in seen:
+                st["drop_dup"] += 1; continue
+            seen.add(h)
+            bid.append(f"{name}_{counter[0]}"); counter[0] += 1
+            btext.append(text); bcre.append(cre); btok.append(tok)
+            st["chars"] += len(text); st["tokens"] += tok; st["kept"] += 1
+            if len(btext) >= 2000:
+                flush()
+    flush(); writer.close(); proc.stdout.close(); proc.wait()
+    st["secs"] = round(time.time() - t0, 1)
+    print(f"  {name}: read {st['read']:,} kept {st['kept']:,} | -short {st['drop_short']:,} "
+          f"-lang {st['drop_lang']:,} -dup {st['drop_dup']:,} -ocr {st['drop_ocr']:,} | "
+          f"{st['chars']/1e6:.0f}M chars, {st['tokens']/1e6:.1f}M tok | {st['secs']}s", flush=True)
+    (out_dir / f"{name}.stats.json").write_text(json.dumps({**st, "license": cfg["license"]}, indent=2))
+    return st
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--sources", nargs="*", default=None)
+    ap.add_argument("--all", action="store_true")
+    ap.add_argument("--speakleash-dir", default="~/speakleash")
+    ap.add_argument("--out", default=".")
+    ap.add_argument("--jobs", type=int, default=os.cpu_count())
+    args = ap.parse_args()
+
+    names = list(SOURCES) if args.all else (args.sources or [])
+    if not names:
+        print("specify --sources <names> or --all"); return
+    sl_dir = Path(args.speakleash_dir).expanduser().resolve()
+    out_root = Path(args.out).expanduser().resolve()
+    print(f"jobs={args.jobs} | speakleash={sl_dir} | out={out_root} | sources={names}", flush=True)
+
+    seen, counter, totals = set(), [0], []
+    t0 = time.time()
+    with mp.Pool(args.jobs, initializer=_init_worker) as pool:
+        for name in names:
+            if name not in SOURCES:
+                print(f"  ? unknown {name}"); continue
+            print(f"[{name}]", flush=True)
+            st = build_source(name, SOURCES[name], sl_dir, out_root, pool, seen, counter)
+            if st:
+                totals.append((name, st))
+    tt = sum(s["tokens"] for _, s in totals)
+    td = sum(s["kept"] for _, s in totals)
+    print(f"\nTOTAL: {td:,} docs, {tt/1e9:.2f}B tok (tiktoken proxy), "
+          f"{len(seen):,} unique | wall {round(time.time()-t0,1)}s", flush=True)
+
+
+if __name__ == "__main__":
+    main()

src/compare_tokenizers.py

@@ -0,0 +1,46 @@
+#!/usr/bin/env python3
+"""Compare fertility of our Polish BPE vs Bielik / Llama-3 / GPT-2 on the same
+held-out Polish sample."""
+import pyarrow.parquet as pq
+from tokenizers import Tokenizer
+import tiktoken
+
+DATA = "/home/ubuntu/dynaword/data"
+sample = []
+for b in pq.ParquetFile(f"{DATA}/wikipedia/wikipedia.parquet").iter_batches(columns=["text"], batch_size=1000):
+    for x in b.column("text"):
+        sample.append(x.as_py())
+    if len(sample) >= 6000:
+        break
+sample = sample[4000:6000]
+words = sum(len(s.split()) for s in sample)
+chars = sum(len(s) for s in sample)
+print(f"held-out: {len(sample)} docs, {words:,} words, {chars:,} chars\n")
+
+rows = []
+
+def add(name, vocab, ntok):
+    rows.append((name, vocab, ntok / words, ntok / chars))
+
+ours = Tokenizer.from_file("/home/ubuntu/dynaword/polish_bpe_32k.json")
+add("polish-32k (ours)", ours.get_vocab_size(),
+    sum(len(e.ids) for e in ours.encode_batch(sample)))
+
+g2 = tiktoken.get_encoding("gpt2")
+add("gpt2-50k", 50257, sum(len(x) for x in g2.encode_ordinary_batch(sample)))
+
+from transformers import AutoTokenizer
+for label, repo in [("Bielik-11B-v3", "speakleash/Bielik-11B-v3.0-Instruct"),
+                    ("Llama-3", "NousResearch/Meta-Llama-3-8B")]:
+    try:
+        t = AutoTokenizer.from_pretrained(repo)
+        enc = t(sample, add_special_tokens=False)["input_ids"]
+        add(label, t.vocab_size, sum(len(x) for x in enc))
+    except Exception as e:
+        print(f"  {label}: FAILED {str(e)[:140]}")
+
+rows.sort(key=lambda r: r[2])
+base = next(r[2] for r in rows if r[0].startswith("polish-32k"))
+print(f"{'tokenizer':<22}{'vocab':>8}{'tok/word':>11}{'tok/char':>11}{'vs ours':>10}")
+for name, vocab, tpw, tpc in rows:
+    print(f"{name:<22}{vocab:>8}{tpw:>11.3f}{tpc:>11.3f}{tpw/base:>9.2f}x")

src/count_vocab.py

@@ -0,0 +1,32 @@
+"""Full-corpus unique word count. Streams parquet batches into a set.
+
+ponytail: lowercased surface forms (Polish is heavily inflected, so this is
+'distinct word forms', not lemmas). Counts total words too.
+"""
+import re
+import pyarrow.parquet as pq
+
+SOURCES = ["eurlex", "parliamentary", "wikisource", "wikipedia",
+           "dziennik_ustaw", "wolne_lektury"]
+WORD = re.compile(r"[^\W\d_]+", re.UNICODE)  # unicode letters, no digits/punct
+
+total_vocab = set()
+total_words = 0
+print(f"{'source':<16} {'words':>14} {'uniq forms':>13}")
+for s in SOURCES:
+    pf = pq.ParquetFile(f"data/{s}/{s}.parquet")
+    src_vocab = set()
+    src_words = 0
+    for batch in pf.iter_batches(columns=["text"], batch_size=2000):
+        for t in batch.column("text").to_pylist():
+            if not t:
+                continue
+            w = WORD.findall(t.lower())
+            src_words += len(w)
+            src_vocab.update(w)
+    total_words += src_words
+    total_vocab |= src_vocab
+    print(f"{s:<16} {src_words:>14,} {len(src_vocab):>13,}", flush=True)
+
+print(f"\n{'TOTAL (deduped)':<16} {total_words:>14,} {len(total_vocab):>13,}")
+print(f"type/token ratio: {len(total_vocab)/total_words:.5f}")

src/eval_bpb.py

@@ -0,0 +1,57 @@
+#!/usr/bin/env python3
+"""Convert modded-nanogpt's validation loss into paper-ready, tokenizer-agnostic
+metrics: perplexity and bits-per-byte (bpb).
+
+Perplexity depends on the tokenizer, so it is NOT comparable across models with
+different vocabularies. Bits-per-byte normalises by raw UTF-8 bytes and IS
+comparable (this is what to report against Bielik / Llama-based models).
+
+  bpb = val_loss[nats/token] / ln(2) * (tokens / bytes)
+
+Usage:
+  python3 src/eval_bpb.py --val-loss 3.21   # val loss from the training log
+"""
+import argparse, struct
+import numpy as np
+from tokenizers import Tokenizer
+
+VAL = "/home/ubuntu/dynaword/shards/polish_val_000000.bin"
+TOK = "/home/ubuntu/dynaword/polish_bpe_32k.json"
+
+def load_shard(path):
+    with open(path, "rb") as f:
+        header = np.frombuffer(f.read(256 * 4), dtype=np.int32)
+        assert header[0] == 20240520 and header[1] == 1, "bad shard header"
+        ntok = int(header[2])
+        toks = np.frombuffer(f.read(ntok * 2), dtype=np.uint16)
+    return toks
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--val-loss", type=float, required=True, help="nats/token from training log")
+    ap.add_argument("--vocab", type=int, default=32768)
+    args = ap.parse_args()
+
+    toks = load_shard(VAL)
+    tok = Tokenizer.from_file(TOK)
+    # decode in chunks -> UTF-8 bytes (held-out reconstructs to the original text)
+    nbytes = 0
+    step = 1_000_000
+    for i in range(0, len(toks), step):
+        nbytes += len(tok.decode(toks[i:i+step].tolist()).encode("utf-8"))
+    ntok = len(toks)
+
+    ln2 = np.log(2)
+    ppl = float(np.exp(args.val_loss))
+    bpb = args.val_loss / ln2 * (ntok / nbytes)
+    bpt = nbytes / ntok
+    rand_bpb = np.log(args.vocab) / ln2 * (ntok / nbytes)   # uniform baseline
+
+    print(f"held-out val: {ntok:,} tokens | {nbytes:,} bytes | {bpt:.3f} bytes/token")
+    print(f"val loss     : {args.val_loss:.4f} nats/token")
+    print(f"perplexity   : {ppl:.2f}   (tokenizer-specific; NOT cross-model comparable)")
+    print(f"bits-per-byte: {bpb:.4f}   (tokenizer-AGNOSTIC; report this)")
+    print(f"  (uniform-{args.vocab} baseline bpb = {rand_bpb:.3f})")
+
+if __name__ == "__main__":
+    main()

src/fetch_eli.py

@@ -0,0 +1,154 @@
+#!/usr/bin/env python3
+"""Fetch Polish primary legislation from the Sejm ELI API into a SpeakLeash-style
+.jsonl.zst, so build_dynaword.py consumes it unchanged (native source, not via
+SpeakLeash). Acts of law are public domain by statute (art. 4 ustawy o prawie
+autorskim — normative acts are not subject to copyright).
+
+API: https://api.sejm.gov.pl/eli/acts/{DU|MP}/{year}            -> list
+     https://api.sejm.gov.pl/eli/acts/{pub}/{year}/{pos}/text.html -> text
+
+Resumable: appends to <out>/<key>.jsonl, skips addresses already present, then
+compresses to <key>.jsonl.zst at the end of a run.
+
+Usage:
+  python3 src/fetch_eli.py --out ~/Local/Ventures/Slayer/data/speakleash \
+      --publishers DU MP --years 1990-2025 --in-force-only --workers 8
+"""
+from __future__ import annotations
+import argparse, json, re, subprocess, sys, time
+from concurrent.futures import ThreadPoolExecutor
+from html.parser import HTMLParser
+from pathlib import Path
+from urllib.request import urlopen, Request
+from urllib.error import URLError, HTTPError
+
+API = "https://api.sejm.gov.pl/eli/acts"
+KEY = "dziennik_ustaw"  # file stem; build_dynaword reads <KEY>.jsonl.zst
+UA = {"User-Agent": "polish-dynaword/0.1 (+research; openly-licensed corpus)"}
+
+
+def _get(url, tries=4, raw=False):
+    for i in range(tries):
+        try:
+            with urlopen(Request(url, headers=UA), timeout=30) as r:
+                data = r.read()
+                return data if raw else json.loads(data)
+        except (HTTPError, URLError, TimeoutError) as e:
+            if isinstance(e, HTTPError) and e.code == 404:
+                return None
+            time.sleep(1.5 * (i + 1))  # ponytail: linear backoff, fine at this scale
+    return None
+
+
+class _Text(HTMLParser):
+    _SKIP = {"script", "style", "head"}
+
+    def __init__(self):
+        super().__init__()
+        self.buf, self._skip = [], 0
+
+    def handle_starttag(self, tag, attrs):
+        if tag in self._SKIP:
+            self._skip += 1
+
+    def handle_endtag(self, tag):
+        if tag in self._SKIP and self._skip:
+            self._skip -= 1
+
+    def handle_data(self, data):
+        if not self._skip and data.strip():
+            self.buf.append(data)
+
+
+def html_to_text(html_bytes):
+    p = _Text()
+    p.feed(html_bytes.decode("utf-8", "replace"))
+    return re.sub(r"\n{3,}", "\n\n", re.sub(r"[ \t]+", " ", "\n".join(p.buf))).strip()
+
+
+def list_acts(pub, year, in_force_only):
+    d = _get(f"{API}/{pub}/{year}")
+    items = (d or {}).get("items", []) if isinstance(d, dict) else []
+    out = []
+    for it in items:
+        if not it.get("textHTML"):
+            continue
+        if in_force_only and it.get("inForce") not in (None, "IN_FORCE"):
+            continue
+        out.append(it)
+    return out
+
+
+def fetch_one(it):
+    pub, year, pos = it["publisher"], it["address"], it["pos"]
+    yr = it.get("year") or int(re.search(r"(\d{4})", it["address"]).group(1))
+    html = _get(f"{API}/{it['publisher']}/{yr}/{it['pos']}/text.html", raw=True)
+    if not html:
+        return None
+    text = html_to_text(html)
+    if len(text) < 200:
+        return None
+    return {
+        "text": text,
+        "meta": {
+            "url": f"{API}/{it['publisher']}/{yr}/{it['pos']}",
+            "title": it.get("title", ""),
+            "address": it["address"], "publisher": it["publisher"],
+            "year": yr, "pos": it["pos"], "type": it.get("type", ""),
+            "status": it.get("status", ""),
+            "announcementDate": it.get("announcementDate", ""),
+        },
+    }
+
+
+def parse_years(s):
+    if "-" in s:
+        a, b = s.split("-"); return range(int(a), int(b) + 1)
+    return [int(s)]
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--out", default="~/Local/Ventures/Slayer/data/speakleash")
+    ap.add_argument("--publishers", nargs="+", default=["DU", "MP"])
+    ap.add_argument("--years", default="1990-2025")
+    ap.add_argument("--in-force-only", action="store_true")
+    ap.add_argument("--workers", type=int, default=8)
+    args = ap.parse_args()
+
+    out_dir = Path(args.out).expanduser(); out_dir.mkdir(parents=True, exist_ok=True)
+    jsonl = out_dir / f"{KEY}.jsonl"
+    done = set()
+    if jsonl.exists():
+        for ln in jsonl.open(encoding="utf-8"):
+            try: done.add(json.loads(ln)["meta"]["address"])
+            except Exception: pass
+        print(f"resume: {len(done):,} already fetched")
+
+    years = parse_years(args.years)
+    todo = []
+    for pub in args.publishers:
+        for y in years:
+            for it in list_acts(pub, y, args.in_force_only):
+                if it["address"] not in done:
+                    todo.append(it)
+            print(f"  listed {pub} {y} | queue={len(todo):,}", flush=True)
+    print(f"to fetch: {len(todo):,}", flush=True)
+
+    kept = 0; t0 = time.time()
+    with jsonl.open("a", encoding="utf-8") as fo, \
+         ThreadPoolExecutor(max_workers=args.workers) as ex:
+        for i, rec in enumerate(ex.map(fetch_one, todo), 1):
+            if rec:
+                fo.write(json.dumps(rec, ensure_ascii=False) + "\n"); kept += 1
+            if i % 500 == 0:
+                print(f"  {i:,}/{len(todo):,} | kept {kept:,} | {i/(time.time()-t0):.1f}/s", flush=True)
+
+    print(f"fetched {kept:,} new | total file lines below; compressing...", flush=True)
+    subprocess.run(["zstd", "-19", "-f", "--rm", str(jsonl), "-o",
+                    str(out_dir / f"{KEY}.jsonl.zst")], check=True)
+    print(f"wrote {out_dir / (KEY + '.jsonl.zst')} in {round(time.time()-t0)}s")
+
+
+if __name__ == "__main__":
+    main()

src/fetch_eltec.py

@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+"""Fetch the Polish ELTeC corpus (European Literary Text Collection, COST Action
+'Distant Reading') into a SpeakLeash-style .jsonl.zst. ~100 Polish novels
+(1840-1920), level-1 TEI XML, released CC-BY. Directly sourced from GitHub.
+
+Usage:
+  python3 src/fetch_eltec.py --out ~/Local/Ventures/Slayer/data/speakleash
+"""
+from __future__ import annotations
+import argparse, json, re, subprocess, time
+from pathlib import Path
+from urllib.request import urlopen, Request
+
+API = "https://api.github.com/repos/COST-ELTeC/ELTeC-pol/contents/level1"
+TEI = "{http://www.tei-c.org/ns/1.0}"
+UA = {"User-Agent": "polish-dynaword/0.1 (+research; openly-licensed corpus)"}
+KEY = "eltec_pol"
+MIN_CHARS = 200
+
+
+def _get(url, raw=False, tries=4):
+    for i in range(tries):
+        try:
+            with urlopen(Request(url, headers=UA), timeout=60) as r:
+                data = r.read()
+                return data if raw else json.loads(data)
+        except Exception:
+            time.sleep(1.5 * (i + 1))
+    return None
+
+
+def tei_to_text(xml_bytes):
+    """Extract reading text (title + body paragraphs) from a TEI file."""
+    from xml.etree.ElementTree import fromstring
+    root = fromstring(xml_bytes)
+    title_el = root.find(f".//{TEI}titleStmt/{TEI}title")
+    title = (title_el.text or "").strip() if title_el is not None else ""
+    body = root.find(f".//{TEI}text/{TEI}body")
+    if body is None:
+        return title, ""
+    parts = []
+    for p in body.iter(f"{TEI}p"):
+        txt = " ".join(p.itertext())
+        txt = re.sub(r"\s+", " ", txt).strip()
+        if txt:
+            parts.append(txt)
+    return title, "\n\n".join(parts)
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--out", default="~/Local/Ventures/Slayer/data/speakleash")
+    args = ap.parse_args()
+    out_dir = Path(args.out).expanduser(); out_dir.mkdir(parents=True, exist_ok=True)
+
+    listing = _get(API) or []
+    files = [f for f in listing if f["name"].endswith(".xml")]
+    print(f"ELTeC-pol: {len(files)} TEI files", flush=True)
+
+    jsonl = out_dir / f"{KEY}.jsonl"; kept = 0; t0 = time.time()
+    with jsonl.open("w", encoding="utf-8") as fo:
+        for i, f in enumerate(files, 1):
+            xml = _get(f["download_url"], raw=True)
+            if not xml:
+                print(f"  ! skip {f['name']}"); continue
+            title, text = tei_to_text(xml)
+            if len(text) < MIN_CHARS:
+                continue
+            fo.write(json.dumps({
+                "text": text,
+                "meta": {"url": f["html_url"], "title": title,
+                         "file": f["name"], "domain": "literature"},
+            }, ensure_ascii=False) + "\n")
+            kept += 1
+            if i % 20 == 0:
+                print(f"  {i}/{len(files)} | kept {kept}", flush=True)
+    subprocess.run(["zstd", "-19", "-f", "--rm", str(jsonl),
+                    "-o", str(out_dir / f"{KEY}.jsonl.zst")], check=True)
+    print(f"{KEY}: {kept} novels in {round(time.time()-t0)}s -> {KEY}.jsonl.zst")
+
+
+if __name__ == "__main__":
+    main()

src/fetch_wikimedia.py

@@ -0,0 +1,105 @@
+#!/usr/bin/env python3
+"""Fetch Polish Wikimedia sibling projects (Wikinews/Wikivoyage/Wikibooks/Wikiquote)
+from the official dumps into SpeakLeash-style .jsonl.zst, so build_dynaword.py
+consumes them unchanged. All are CC-BY-SA (same basis as Wikipedia/Wikisource).
+
+Mainspace (ns=0) only; redirects and empty pages dropped. Wikitext -> plaintext
+via mwparserfromhell.
+
+Usage:
+  python3 src/fetch_wikimedia.py --out ~/Local/Ventures/Slayer/data/speakleash \
+      --wikis plwikinews plwikivoyage plwikibooks plwikiquote
+"""
+from __future__ import annotations
+import argparse, bz2, json, subprocess, time
+from pathlib import Path
+from urllib.request import urlopen, Request
+from xml.etree.ElementTree import iterparse
+import mwparserfromhell
+
+DUMP = "https://dumps.wikimedia.org/{w}/latest/{w}-latest-pages-articles.xml.bz2"
+UA = {"User-Agent": "polish-dynaword/0.1 (+research; openly-licensed corpus)"}
+MIN_CHARS = 200
+# wiki -> (file stem used by build_dynaword, human URL)
+SITE = {
+    "plwikinews":   ("https://pl.wikinews.org/wiki/", "news"),
+    "plwikivoyage": ("https://pl.wikivoyage.org/wiki/", "travel"),
+    "plwikibooks":  ("https://pl.wikibooks.org/wiki/", "howto"),
+    "plwikiquote":  ("https://pl.wikiquote.org/wiki/", "quotes"),
+}
+
+
+def _local(tag):  # strip {namespace}
+    return tag.rsplit("}", 1)[-1]
+
+
+def download(w, cache):
+    dst = cache / f"{w}.xml.bz2"
+    if dst.exists() and dst.stat().st_size > 0:
+        print(f"  cached {dst.name}"); return dst
+    url = DUMP.format(w=w)
+    print(f"  downloading {url}", flush=True)
+    with urlopen(Request(url, headers=UA), timeout=120) as r, dst.open("wb") as f:
+        while chunk := r.read(1 << 20):
+            f.write(chunk)
+    return dst
+
+
+def pages(bz2_path):
+    """Yield (title, wikitext) for mainspace, non-redirect pages."""
+    with bz2.open(bz2_path, "rb") as fh:
+        title = ns = text = None; redirect = False
+        for ev, el in iterparse(fh, events=("end",)):
+            t = _local(el.tag)
+            if t == "title": title = el.text
+            elif t == "ns": ns = el.text
+            elif t == "redirect": redirect = True
+            elif t == "text": text = el.text
+            elif t == "page":
+                if ns == "0" and not redirect and text:
+                    yield title, text
+                title = ns = text = None; redirect = False
+                el.clear()
+
+
+def strip(wikitext):
+    try:
+        return mwparserfromhell.parse(wikitext).strip_code(normalize=True, collapse=True).strip()
+    except Exception:
+        return ""
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--out", default="~/Local/Ventures/Slayer/data/speakleash")
+    ap.add_argument("--wikis", nargs="+", default=list(SITE))
+    args = ap.parse_args()
+    out_dir = Path(args.out).expanduser(); out_dir.mkdir(parents=True, exist_ok=True)
+    cache = out_dir / "_wikimedia_dumps"; cache.mkdir(exist_ok=True)
+
+    for w in args.wikis:
+        base_url, domain = SITE[w]
+        print(f"[{w}] ({domain})", flush=True)
+        src = download(w, cache)
+        jsonl = out_dir / f"{w}.jsonl"
+        kept = 0; t0 = time.time()
+        with jsonl.open("w", encoding="utf-8") as fo:
+            for title, wt in pages(src):
+                text = strip(wt)
+                if len(text) < MIN_CHARS:
+                    continue
+                fo.write(json.dumps({
+                    "text": text,
+                    "meta": {"url": base_url + (title or "").replace(" ", "_"),
+                             "title": title, "wiki": w, "domain": domain},
+                }, ensure_ascii=False) + "\n")
+                kept += 1
+                if kept % 2000 == 0:
+                    print(f"  {kept:,} kept | {kept/(time.time()-t0):.0f}/s", flush=True)
+        subprocess.run(["zstd", "-19", "-f", "--rm", str(jsonl),
+                        "-o", str(out_dir / f"{w}.jsonl.zst")], check=True)
+        print(f"  {w}: {kept:,} docs in {round(time.time()-t0)}s -> {w}.jsonl.zst", flush=True)
+
+
+if __name__ == "__main__":
+    main()

src/make_docs.py

@@ -0,0 +1,219 @@
+#!/usr/bin/env python3
+"""Generate Dynaword documentation: per-source datasheets + README + CHANGELOG + LICENSE.
+
+Reads sources.py + data/<source>/<source>.stats.json (written by build_dynaword.py).
+Implements the "Documented" principle (datasheets, Gebru et al. 2021) and the
+aggregate README table (paper 2508.02271).
+"""
+from __future__ import annotations
+import json, sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parent))
+from sources import SOURCES, EXCLUDED, ADDED
+
+ROOT = Path(__file__).resolve().parent.parent
+VERSION = "0.2.0"
+CONTACT = "k.wikiel@gmail.com"  # notice-and-takedown / data-removal requests
+
+# Dupochron: documented good-faith provenance + no-warranty + takedown + PII.
+DISCLAIMER = f"""## Personal & sensitive data
+This corpus contains **only** text that its upstream sources already published
+under open licenses or as official public-domain record. It therefore includes
+names and statements of **public figures acting in a public capacity** — e.g.
+parliamentary speakers (PPC), authorities named in legal acts (EUR-Lex), and
+people described in encyclopedic articles (Wikipedia/Wikisource). No private,
+non-public personal data was collected or added. If you are a data subject and
+want content concerning you removed, contact **{CONTACT}** — it will be dropped
+from the next version (see retroactive-removal policy below).
+
+## Disclaimer & legal
+- **Provenance in good faith.** Per-source licenses are reproduced *as documented
+  by the upstream sources and by SpeakLeash* (the intermediate aggregator), to the
+  best of our knowledge. We make no independent legal warranty about the copyright
+  status of any individual document.
+- **No ownership claim.** This release is a *curated, license-reviewed, documented
+  aggregation*. We claim no ownership of the underlying texts; rights remain with
+  the original authors/rightsholders under their respective licenses.
+- **Provided "as is"**, without warranty of any kind, express or implied. This is
+  not legal advice.
+- **Your compliance is yours.** Downstream users must satisfy each upstream
+  license themselves — in particular **CC-BY-SA-4.0 attribution and share-alike**
+  for derivatives of this dataset, and attribution to the upstream sources and to
+  SpeakLeash.
+- **Notice-and-takedown.** Any source or rightsholder raising a substantiated
+  objection can have material removed: contact **{CONTACT}**; it is dropped from
+  the next version and recorded in the CHANGELOG. Removal is retroactive
+  going-forward (prior immutable snapshots/commits may persist).
+"""
+
+
+def load_stats(name):
+    f = ROOT / "data" / name / f"{name}.stats.json"
+    return json.loads(f.read_text()) if f.exists() else None
+
+
+def datasheet(name, cfg, st):
+    return f"""# {name}
+
+{cfg['pretty']}
+
+## Dataset description
+- **Source (upstream):** {cfg['upstream']}
+- **Domain:** {cfg['domain']}
+- **Language:** Polish (pl)
+- **License:** `{cfg['license']}`
+- **Created (range):** {cfg['created']}
+- **Added:** {ADDED}
+
+## Licensing — traceable basis
+{cfg['traceable']}
+
+## Provenance
+{cfg.get('provenance', f"Pulled from SpeakLeash's public redistribution "
+ f"(`speakleash-ds-pub`, key `{cfg.get('speakleash_key')}`) of the upstream source "
+ f"above. SpeakLeash credited as intermediate aggregator; upstream "
+ f"license/attribution preserved.")}
+
+## Statistics
+| documents | characters | tokens (tiktoken proxy) |
+|---:|---:|---:|
+| {st['kept']:,} | {st['chars']:,} | {st['tokens']:,} |
+
+## Filters applied (build_dynaword.py)
+Minimal, per Dynaword guidelines (heavy filtering left to downstream use):
+- drop documents < 200 chars: **{st['drop_short']:,}**
+- drop non-Polish (diacritic ratio): **{st['drop_lang']:,}**
+- exact cross-source dedup (sha1): **{st['drop_dup']:,}**
+- OCR alpha-ratio < 0.70 (OCR sources only): **{st['drop_ocr']:,}**
+- read {st['read']:,} → kept {st['kept']:,}
+
+Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
+Llama-3 count is computed at release.
+"""
+
+
+def main():
+    rows, tot_doc, tot_tok, tot_chr = [], 0, 0, 0
+    for name, cfg in SOURCES.items():
+        st = load_stats(name)
+        if not st:
+            print(f"  ! no stats for {name}"); continue
+        (ROOT / "data" / name / f"{name}.md").write_text(datasheet(name, cfg, st))
+        rows.append((name, cfg, st))
+        tot_doc += st["kept"]; tot_tok += st["tokens"]; tot_chr += st["chars"]
+    rows.sort(key=lambda r: -r[2]["tokens"])
+
+    tbl = "\n".join(
+        f"| [{n}](data/{n}/{n}.md) | {c['pretty']} | `{c['license']}` | "
+        f"{s['kept']:,} | {s['tokens']/1e6:,.1f}M |"
+        for n, c, s in rows)
+    excl = "\n".join(f"| `{k}` | {v} |" for k, v in EXCLUDED.items())
+
+    readme = f"""---
+license: cc-by-sa-4.0
+language:
+- pl
+pretty_name: Polish DynaWord
+task_categories:
+- text-generation
+size_categories:
+- 1M<n<10M
+tags:
+- polish
+- pretraining
+- dynaword
+---
+
+# Polish DynaWord
+
+A continuously developed, **openly-licensed**, human-text Polish corpus — a Polish
+edition in the [Dynaword](https://huggingface.co/datasets/danish-foundation-models/danish-dynaword)
+family (Enevoldsen et al., [arXiv:2508.02271](https://arxiv.org/abs/2508.02271)).
+
+> **v{VERSION}** · {tot_doc:,} documents · **{tot_tok/1e9:.2f}B tokens** (tiktoken proxy;
+> canonical Llama-3 count at release) · {len(rows)} sources
+
+## What this dataset contributes
+The raw texts come from existing open corpora (redistributed via SpeakLeash and,
+where applicable, fetched from upstream). **The value added here is the curation,
+not the bytes**, following the Dynaword methodology:
+
+1. **License review per source** — each source vetted for an *openly-licensed,
+   traceable* legal basis (documented in its datasheet); sources that fail the
+   review are **excluded with a stated reason** (see table below), not silently
+   kept. This is the core editorial work.
+2. **Filtering & normalization** — minimal, reproducible gates (short-doc,
+   non-Polish, exact cross-source dedup, OCR garble) applied uniformly to one
+   clean schema: `id, text, source, added, created, token_count`.
+3. **Documentation** — a datasheet per source (Gebru et al. 2021) + this card,
+   so provenance and licensing are auditable rather than assumed.
+4. **Reproducibility & versioning** — `src/` rebuilds the corpus from sources;
+   new sources and removals are tracked in the CHANGELOG.
+
+Credit for the underlying texts belongs to the upstream sources and to SpeakLeash
+as the redistributing aggregator; this release does not claim ownership of them
+(see Disclaimer).
+
+## Guiding principles
+1. **Open & traceable licensing** — every source is *openly licensed* with a documented
+   legal basis (see each datasheet's "traceable basis"), not a vague "public domain".
+2. **Reproducibility** — `src/build_dynaword.py` rebuilds the corpus from sources.
+3. **Documented** — a datasheet per source under `data/<source>/`.
+4. **Extensibility** — versioned; new sources via PR.
+
+## Sources
+| source | description | license | documents | tokens |
+|---|---|---|---:|---:|
+{tbl}
+| **total** | | | **{tot_doc:,}** | **{tot_tok/1e6:,.1f}M** |
+
+## Method
+Only **human-authored** text — no synthetic, machine-translated, or auto-transcribed
+data. Gates are intentionally minimal (drop short docs, non-Polish, exact duplicates,
+OCR garble); heavy quality filtering and mix-weighting are left to downstream training.
+Evaluation-set decontamination is applied/marked separately. Schema:
+`id, text, source, added, created, token_count`.
+
+## Excluded sources (transparency)
+Sources we reviewed and **deliberately left out** — part of the curation:
+
+| source | reason |
+|---|---|
+{excl}
+
+{DISCLAIMER}
+## License & attribution
+Released under **CC-BY-SA-4.0** (copyleft inherited from CC-BY-SA sources such as
+Wikipedia/Wikisource/Wolne Lektury). Attribution due to each upstream (see datasheets)
+and to **SpeakLeash** as the intermediate aggregator. Retroactive-removal policy: a
+source that raises an objection is dropped from subsequent versions, recorded in the
+CHANGELOG.
+
+## Reproduce
+```bash
+python3 src/build_dynaword.py --all --speakleash-dir <speakleash_zst_dir> --out .
+python3 src/make_docs.py
+```
+"""
+    (ROOT / "README.md").write_text(readme)
+
+    (ROOT / "CHANGELOG.md").write_text(
+        f"# Changelog\n\n## v{VERSION} ({ADDED})\n\n"
+        f"- Initial release: {len(rows)} openly-licensed sources, "
+        f"{tot_doc:,} docs, {tot_tok/1e9:.2f}B tokens (tiktoken proxy).\n"
+        f"- Sources: {', '.join(n for n, _, _ in rows)}.\n"
+        f"- Excluded (see README): {', '.join(EXCLUDED)}.\n")
+
+    (ROOT / "LICENSE").write_text(
+        "Polish DynaWord is released under Creative Commons Attribution-ShareAlike 4.0\n"
+        "International (CC-BY-SA-4.0): https://creativecommons.org/licenses/by-sa/4.0/\n\n"
+        "Per-source upstream licenses and attribution are documented in each\n"
+        "data/<source>/<source>.md datasheet.\n")
+
+    print(f"docs written: README + CHANGELOG + LICENSE + {len(rows)} datasheets")
+    print(f"TOTAL {tot_doc:,} docs | {tot_tok/1e9:.2f}B tok | {tot_chr/1e9:.1f}B chars")
+
+
+if __name__ == "__main__":
+    main()

src/make_linkedin_card.py

@@ -0,0 +1,98 @@
+#!/usr/bin/env python3
+"""Beautiful LinkedIn teaser card for the open-licensed Polish token dataset.
+Vague on provenance: shows totals + domain split + licensing claims only.
+"""
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+from matplotlib.patches import FancyBboxPatch
+from matplotlib import font_manager as fm
+
+FT = "Avenir Next" if "Avenir Next" in {f.name for f in fm.fontManager.ttflist} else "Helvetica Neue"
+plt.rcParams["font.family"] = FT
+
+# palette
+BG     = "#0B0F14"
+INK    = "#EAF1F8"
+MUTED  = "#7E8B9A"
+FAINT  = "#1B232D"
+GREEN  = "#34E0A1"   # legal & official
+BLUE   = "#5AA2FF"   # encyclopedic
+VIOLET = "#B98BFF"   # literature
+
+# data (vague domains; numbers in millions of tokens, tiktoken proxy)
+domains = [("Legal & official", 4024.9, GREEN),
+           ("Encyclopedic & reference", 1509.1, BLUE),
+           ("Literature & books", 104.0, VIOLET)]
+total = sum(v for _, v, _ in domains)
+
+fig = plt.figure(figsize=(12, 12), dpi=100)
+fig.patch.set_facecolor(BG)
+ax = fig.add_axes([0, 0, 1, 1]); ax.set_xlim(0, 1); ax.set_ylim(0, 1); ax.axis("off")
+ax.add_patch(plt.Rectangle((0, 0), 1, 1, color=BG, zorder=-10))
+
+# top accent bar
+ax.add_patch(plt.Rectangle((0, 0.988), 1, 0.012, color=GREEN, zorder=5))
+
+# eyebrow
+ax.text(0.07, 0.915, "P O L I S H   L A N G U A G E   ·   O P E N   P R E T R A I N I N G   D A T A",
+        color=GREEN, fontsize=14.5, weight="bold", alpha=0.95)
+# hero
+ax.text(0.07, 0.85, "Licensed to train.", color=INK, fontsize=52, weight="bold")
+ax.text(0.072, 0.788,
+        "A Polish corpus where every single token is openly licensed —\ncleared for model training, with a documented legal basis.",
+        color=MUTED, fontsize=18.5, linespacing=1.45)
+
+# donut (left)
+dax = fig.add_axes([0.045, 0.30, 0.44, 0.44]); dax.set_aspect("equal")
+vals = [v for _, v, _ in domains]; cols = [c for _, _, c in domains]
+dax.pie(vals, colors=cols, startangle=90, counterclock=False,
+        wedgeprops=dict(width=0.30, edgecolor=BG, linewidth=4))
+dax.text(0, 0.10, "5.6B", ha="center", va="center", color=INK, fontsize=58, weight="bold")
+dax.text(0, -0.12, "openly-licensed tokens", ha="center", va="center", color=MUTED, fontsize=16)
+
+# legend (right)
+lx, ly = 0.565, 0.625
+for name, v, c in domains:
+    ax.add_patch(plt.Circle((lx, ly + 0.011), 0.013, color=c, transform=ax.transAxes))
+    ax.text(lx + 0.035, ly + 0.020, name, color=INK, fontsize=20, weight="bold", va="center")
+    ax.text(lx + 0.035, ly - 0.012, f"{v/total*100:.0f}%   ·   {v/1000:.1f}B tokens",
+            color=MUTED, fontsize=14.5, va="center")
+    ly -= 0.085
+
+# stat strip
+ax.add_patch(plt.Rectangle((0.07, 0.247), 0.86, 0.0016, color=FAINT))
+stats = [("2.4M", "documents"), ("100%", "open licenses"),
+         ("0", "copyrighted media"), ("human", "written, not synthetic")]
+sx = 0.07
+for big, small in stats:
+    ax.text(sx, 0.205, big, color=INK, fontsize=27, weight="bold")
+    ax.text(sx, 0.168, small, color=MUTED, fontsize=14.5)
+    sx += 0.225
+
+# license pills (centered row, drawn checkmarks since the font lacks U+2713)
+pills = [("Open licenses only", GREEN), ("Traceable provenance", BLUE),
+         ("Deduplicated", VIOLET)]
+def pw(text): return 0.0125 * len(text) + 0.062
+widths = [pw(t) for t, _ in pills]
+gap, ph, py = 0.022, 0.05, 0.088
+total_w = sum(widths) + gap * (len(pills) - 1)
+px = (1 - total_w) / 2
+for (t, c), w in zip(pills, widths):
+    ax.add_patch(FancyBboxPatch((px, py), w, ph, boxstyle="round,pad=0.004,rounding_size=0.024",
+                 mutation_aspect=1, fc=FAINT, ec=c, lw=1.7, zorder=3))
+    cx = px + 0.026
+    ax.plot([cx, cx + 0.008, cx + 0.02], [py + ph/2, py + ph/2 - 0.008, py + ph/2 + 0.012],
+            color=c, lw=2.8, solid_capstyle="round", solid_joinstyle="round", zorder=5)
+    ax.text(px + w/2 + 0.016, py + ph/2, t, ha="center", va="center",
+            color=INK, fontsize=13.5, weight="bold", zorder=4)
+    px += w + gap
+
+# footer
+ax.text(0.07, 0.04, "CC-BY-SA  ·  public domain  ·  permissive-only", color=MUTED, fontsize=14)
+ax.text(0.93, 0.04, "SLAYER  LABS", color=INK, fontsize=15, weight="bold", ha="right", alpha=0.9)
+
+import os
+out = os.path.expanduser("~/Desktop/polish_open_corpus_card.png")
+fig.savefig(out, facecolor=BG, dpi=100)
+print("saved", out)

src/rights.py

@@ -0,0 +1,43 @@
+"""edm:rights classifier for the Europeana source (mixed per-record rights).
+
+Dynaword principle = openly licensed. Allow only PD / CC0 / CC-BY / CC-BY-SA.
+NC/ND are not open; InC / NoC-* / CNE are in-copyright or unclear -> drop.
+Wire into build_dynaword.py only for europeana (per-record); other sources keep
+their single per-source license.
+"""
+
+# ponytail: substring match on the rights URI — finite, well-known statement set
+_OPEN = (
+    "creativecommons.org/publicdomain/mark",   # Public Domain Mark
+    "creativecommons.org/publicdomain/zero",   # CC0
+    "creativecommons.org/licenses/by/",        # CC-BY
+    "creativecommons.org/licenses/by-sa/",     # CC-BY-SA
+)
+
+
+def rights_ok(uri: str) -> bool:
+    u = (uri or "").lower().rstrip("/") + "/"
+    return any(tag in u for tag in _OPEN)
+
+
+def spdx(uri: str) -> str:
+    u = (uri or "").lower()
+    if "publicdomain/zero" in u: return "CC0-1.0"
+    if "publicdomain/mark" in u: return "PD"
+    if "/by-sa/" in u: return "CC-BY-SA-4.0"
+    if "/by/" in u:    return "CC-BY-4.0"
+    return "UNKNOWN"
+
+
+if __name__ == "__main__":  # self-check
+    assert rights_ok("http://creativecommons.org/licenses/by-sa/4.0/")
+    assert rights_ok("https://creativecommons.org/publicdomain/mark/1.0/")
+    assert rights_ok("http://creativecommons.org/publicdomain/zero/1.0/")
+    assert not rights_ok("http://creativecommons.org/licenses/by-nc/4.0/")
+    assert not rights_ok("http://creativecommons.org/licenses/by-nd/4.0/")
+    assert not rights_ok("http://rightsstatements.org/vocab/InC/1.0/")
+    assert not rights_ok("http://rightsstatements.org/vocab/NoC-OKLR/1.0/")
+    assert not rights_ok("")
+    assert spdx("http://creativecommons.org/licenses/by-sa/4.0/") == "CC-BY-SA-4.0"
+    assert spdx("http://creativecommons.org/publicdomain/zero/1.0/") == "CC0-1.0"
+    print("ok")

src/sample_quality.py

@@ -0,0 +1,81 @@
+"""Sample docs per source and score corpus quality with cheap heuristics.
+
+ponytail: heuristic eyeball score, not an LLM judge. Add LLM-judge pass if these
+metrics look ambiguous and you need semantic quality, not just surface garble.
+"""
+import sys, re, random
+import pyarrow.parquet as pq
+
+SOURCES = ["eurlex", "parliamentary", "wikisource", "wikipedia",
+           "dziennik_ustaw", "wolne_lektury"]
+N = int(sys.argv[1]) if len(sys.argv) > 1 else 200  # docs sampled per source
+SHOW = 2  # raw samples printed per source
+
+PL_DIAC = set("ąćęłńóśźżĄĆĘŁŃÓŚŹŻ")
+LETTER = re.compile(r"[^\W\d_]", re.UNICODE)
+
+def sample_rows(path, n):
+    """Grab n docs spread across scattered row groups (cheap pseudo-random)."""
+    pf = pq.ParquetFile(path)
+    ng = pf.num_row_groups
+    groups = sorted(random.sample(range(ng), min(ng, 8)))
+    out = []
+    per = max(1, n // len(groups))
+    for g in groups:
+        tbl = pf.read_row_group(g, columns=["text"])
+        texts = tbl.column("text").to_pylist()
+        out += random.sample(texts, min(per, len(texts)))
+        if len(out) >= n:
+            break
+    return out[:n]
+
+def score(text):
+    t = text or ""
+    n = len(t)
+    if n == 0:
+        return None
+    letters = sum(1 for c in t if LETTER.match(c))
+    diac = sum(1 for c in t if c in PL_DIAC)
+    repl = t.count("�")                      # OCR/decode garbage
+    digits = sum(c.isdigit() for c in t)
+    space = sum(c.isspace() for c in t)
+    words = t.split()
+    uniq = len(set(words)) / len(words) if words else 0
+    return dict(
+        chars=n,
+        letter_ratio=letters / n,
+        diac_per_kchar=1000 * diac / n,
+        repl_per_kchar=1000 * repl / n,
+        digit_ratio=digits / n,
+        space_ratio=space / n,
+        word_uniq=uniq,
+    )
+
+def avg(rows, k):
+    v = [r[k] for r in rows if r]
+    return sum(v) / len(v) if v else 0
+
+print(f"sampling {N} docs/source\n")
+print(f"{'source':<16} {'chars':>8} {'letter%':>8} {'diac/k':>7} "
+      f"{'repl/k':>7} {'digit%':>7} {'uniq':>6} {'<200ch':>7}")
+for s in SOURCES:
+    path = f"data/{s}/{s}.parquet"
+    docs = sample_rows(path, N)
+    scored = [score(d) for d in docs]
+    scored = [x for x in scored if x]
+    short = sum(1 for x in scored if x["chars"] < 200) / len(scored)
+    print(f"{s:<16} {avg(scored,'chars'):>8.0f} "
+          f"{100*avg(scored,'letter_ratio'):>7.1f}% "
+          f"{avg(scored,'diac_per_kchar'):>7.1f} "
+          f"{avg(scored,'repl_per_kchar'):>7.2f} "
+          f"{100*avg(scored,'digit_ratio'):>6.1f}% "
+          f"{avg(scored,'word_uniq'):>6.2f} "
+          f"{100*short:>6.1f}%")
+
+print("\n=== raw samples ===")
+for s in SOURCES:
+    docs = sample_rows(f"data/{s}/{s}.parquet", SHOW)
+    print(f"\n--- {s} ---")
+    for d in docs:
+        snippet = re.sub(r"\s+", " ", d)[:300]
+        print(f"  [{len(d)} ch] {snippet}")

src/sources.py

@@ -0,0 +1,188 @@
+"""Source registry for Polish DynaWord.
+
+Each entry = one openly-licensed, *traceable* human-text source (paper 2508.02271,
+principle: open+traceable licensing). The `traceable` field documents WHY the
+license holds (not just "it's public domain") — e.g. statutory basis or upstream
+license — as required by the Dynaword guidelines.
+
+Provenance note: data is pulled from SpeakLeash's public redistribution
+(iDrive e2 bucket) of these upstream sources; SpeakLeash is credited as the
+intermediate aggregator, upstream license/attribution is preserved per source.
+"""
+
+# clean-6: passed the copyright review (2026-06-15). open_subtitles excluded
+# (derivative of copyrighted film dialogue — same lesson as Danish Gigaword),
+# europeana + 1000_novels deferred pending per-record rights / title check.
+SOURCES = {
+    "wikipedia": {
+        "speakleash_key": "plwiki",
+        "pretty": "Polish Wikipedia",
+        "license": "CC-BY-SA-3.0",
+        "license_spdx": "CC-BY-SA-3.0",
+        "traceable": "Wikimedia dumps explicitly released under CC-BY-SA 3.0; "
+                     "attribution to contributors + share-alike preserved.",
+        "upstream": "https://dumps.wikimedia.org/plwiki/latest/",
+        "domain": "encyclopedic",
+        "created": "2001-01-15, 2026-06-14",
+        "is_ocr": False,
+    },
+    "wikisource": {
+        "speakleash_key": "plwikisource",
+        "pretty": "Polish Wikisource",
+        "license": "CC-BY-SA-3.0",
+        "license_spdx": "CC-BY-SA-3.0",
+        "traceable": "Wikimedia dumps under CC-BY-SA 3.0; underlying texts are "
+                     "public-domain works transcribed by the community.",
+        "upstream": "https://dumps.wikimedia.org/plwikisource/latest/",
+        "domain": "books/source-texts",
+        "created": "2005-01-01, 2026-06-14",
+        "is_ocr": False,
+    },
+    "eurlex": {
+        "speakleash_key": "eurlex_corpus",
+        "pretty": "EUR-Lex (EU legal acts, Polish)",
+        "license": "CC-BY-4.0",
+        "license_spdx": "CC-BY-4.0",
+        "traceable": "EU documents reusable under Commission Decision 2011/833/EU; "
+                     "normative acts fall outside copyright (PL art. 4 pr. aut.). "
+                     "Reuse authorised with source acknowledgement.",
+        "upstream": "https://eur-lex.europa.eu",
+        "domain": "legal",
+        "created": "1952-01-01, 2024-12-31",
+        "is_ocr": False,
+    },
+    "parliamentary": {
+        "speakleash_key": "PPC_corpus",
+        "pretty": "Polish Parliamentary Corpus (Sejm/Senat)",
+        "license": "public-domain (official documents)",
+        "license_spdx": "CC-BY-4.0",
+        "traceable": "Parliamentary proceedings are official documents outside "
+                     "copyright (PL art. 4 pr. aut.); PPC redistributed by IPI PAN "
+                     "under CC-BY-4.0.",
+        "upstream": "http://clip.ipipan.waw.pl/PPC",
+        "domain": "political/spoken",
+        "created": "1991-01-01, 2019-12-31",
+        "is_ocr": False,
+    },
+    "wolne_lektury": {
+        "speakleash_key": "wolne_lektury_corpus",
+        "pretty": "Wolne Lektury (school readings)",
+        "license": "CC-BY-SA-4.0 / Wolna Sztuka 1.3",
+        "license_spdx": "CC-BY-SA-4.0",
+        "traceable": "Wolne Lektury publishes under CC-BY-SA 4.0 or Free Art "
+                     "License; texts are public-domain classics + cleared works.",
+        "upstream": "https://wolnelektury.pl/api/books/",
+        "domain": "literature",
+        "created": "1500-01-01, 2024-12-31",
+        "is_ocr": False,
+    },
+    "wikinews": {
+        "file_key": "plwikinews",  # fetched from Wikimedia dumps, NOT via SpeakLeash
+        "pretty": "Polish Wikinews",
+        "license": "CC-BY-2.5",
+        "license_spdx": "CC-BY-2.5",
+        "traceable": "Wikinews content is released under CC-BY 2.5; attribution to "
+                     "contributors preserved.",
+        "upstream": "https://pl.wikinews.org/",
+        "provenance": "Fetched directly from the official Wikimedia dump "
+                      "(plwikinews-latest-pages-articles) by src/fetch_wikimedia.py; "
+                      "mainspace only, wikitext stripped via mwparserfromhell.",
+        "domain": "news",
+        "created": "2005-01-01, 2025-12-31",
+        "is_ocr": False,
+    },
+    "wikivoyage": {
+        "file_key": "plwikivoyage",
+        "pretty": "Polish Wikivoyage (travel guides)",
+        "license": "CC-BY-SA-3.0",
+        "license_spdx": "CC-BY-SA-3.0",
+        "traceable": "Wikivoyage content is released under CC-BY-SA 3.0; "
+                     "attribution + share-alike preserved.",
+        "upstream": "https://pl.wikivoyage.org/",
+        "provenance": "Fetched directly from the official Wikimedia dump "
+                      "(plwikivoyage-latest-pages-articles) by src/fetch_wikimedia.py; "
+                      "mainspace only, wikitext stripped via mwparserfromhell.",
+        "domain": "travel",
+        "created": "2012-01-01, 2025-12-31",
+        "is_ocr": False,
+    },
+    "wikibooks": {
+        "file_key": "plwikibooks",
+        "pretty": "Polish Wikibooks (open textbooks)",
+        "license": "CC-BY-SA-3.0",
+        "license_spdx": "CC-BY-SA-3.0",
+        "traceable": "Wikibooks content is released under CC-BY-SA 3.0; "
+                     "attribution + share-alike preserved.",
+        "upstream": "https://pl.wikibooks.org/",
+        "provenance": "Fetched directly from the official Wikimedia dump "
+                      "(plwikibooks-latest-pages-articles) by src/fetch_wikimedia.py; "
+                      "mainspace only, wikitext stripped via mwparserfromhell.",
+        "domain": "instructional",
+        "created": "2004-01-01, 2025-12-31",
+        "is_ocr": False,
+    },
+    "wikiquote": {
+        "file_key": "plwikiquote",
+        "pretty": "Polish Wikiquote (quotations)",
+        "license": "CC-BY-SA-3.0",
+        "license_spdx": "CC-BY-SA-3.0",
+        "traceable": "Wikiquote content is released under CC-BY-SA 3.0; "
+                     "attribution + share-alike preserved.",
+        "upstream": "https://pl.wikiquote.org/",
+        "provenance": "Fetched directly from the official Wikimedia dump "
+                      "(plwikiquote-latest-pages-articles) by src/fetch_wikimedia.py; "
+                      "mainspace only, wikitext stripped via mwparserfromhell.",
+        "domain": "quotations",
+        "created": "2005-01-01, 2025-12-31",
+        "is_ocr": False,
+    },
+    "eltec_pol": {
+        "file_key": "eltec_pol",
+        "pretty": "ELTeC-pol (European Literary Text Collection, Polish)",
+        "license": "CC-BY-4.0",
+        "license_spdx": "CC-BY-4.0",
+        "traceable": "ELTeC corpora are released under CC-BY 4.0 by the COST Action "
+                     "'Distant Reading for European Literary History' (CA16204).",
+        "upstream": "https://github.com/COST-ELTeC/ELTeC-pol",
+        "provenance": "Fetched directly from the COST-ELTeC/ELTeC-pol GitHub "
+                      "repository (level1 TEI) by src/fetch_eltec.py; reading text "
+                      "extracted from TEI <body> paragraphs.",
+        "domain": "literature",
+        "created": "1840-01-01, 1920-12-31",
+        "is_ocr": False,
+    },
+    "dziennik_ustaw": {
+        "file_key": "dziennik_ustaw",  # fetched directly, NOT via SpeakLeash
+        "pretty": "Dziennik Ustaw + Monitor Polski (Polish primary legislation)",
+        "license": "public-domain (official documents)",
+        "license_spdx": "CC0-1.0",
+        "traceable": "Normative acts and their official drafts are not subject to "
+                     "copyright (PL art. 4 ustawy o prawie autorskim). Published by "
+                     "the state; reuse unrestricted.",
+        "upstream": "https://api.sejm.gov.pl/eli/acts",
+        "provenance": "Fetched directly from the Sejm ELI API (text.html per act) "
+                      "by src/fetch_eli.py — a native source, not redistributed via "
+                      "SpeakLeash. Full provenance: each act's `meta.url` is its ELI "
+                      "address, resolvable on api.sejm.gov.pl.",
+        "domain": "legal",
+        "created": "1918-01-01, 2025-12-31",
+        "is_ocr": False,
+    },
+}
+
+ADDED = "2026-06-15"
+EXCLUDED = {
+    "open_subtitles_corpus": "Derivative of copyrighted film/TV dialogue; "
+        "OpenSubtitles uploads largely unlicensed. Same copyright lesson as "
+        "Danish Gigaword's OpenSubtitles (paper 2508.02271). Not openly licensed.",
+    "europeana_eu_pl_corpus": "Aggregated items with mixed per-record rights "
+        "(PD / CC-BY-NC / rights-reserved). Needs per-record edm:rights filter "
+        "before any inclusion.",
+    "1000_novels_corpus_CLARIN-PL": "CC-BY-4.0 label, but 'novels' likely include "
+        "in-copyright contemporary works; verify titles/years on CLARIN handle "
+        "11321/312 before inclusion.",
+    "project_gutenberg_pl_corpus": "Only 31 PL books (4.3MB) — PG is ~99% English; "
+        "Polish PD literature already covered by wolne_lektury + wikisource (so "
+        "near-redundant after dedup). Dropped to avoid the PD-in-EU per-work check "
+        "(PG claims PD-in-US only) for negligible token gain.",
+}

src/tokenize_shards.py

@@ -0,0 +1,75 @@
+#!/usr/bin/env python3
+"""Tokenize the corpus with the Polish BPE into modded-nanogpt/llm.c shards.
+Single process: tokenizers' encode_batch parallelises across cores via Rust rayon
+natively — NO mp.Pool (nesting mp.Pool x rayon oversubscribes and is ~4x slower).
+Format: 256 int32 header [magic 20240520, version 1, ntok] + uint16 tokens.
+Docs joined with <|endoftext|>; ~1/200 docs held out for validation."""
+import glob, os, time
+import numpy as np
+import pyarrow.parquet as pq
+from tokenizers import Tokenizer
+
+TOKJSON = "/home/ubuntu/dynaword/polish_bpe_32k.json"
+DATA = "/home/ubuntu/dynaword/data"
+OUT = "/home/ubuntu/dynaword/shards"
+SHARD = 100_000_000
+VAL_EVERY = 200
+CHUNK = 50_000           # docs per encode_batch call
+os.makedirs(OUT, exist_ok=True)
+
+def doc_stream():
+    for f in sorted(glob.glob(f"{DATA}/*/*.parquet")):
+        for b in pq.ParquetFile(f).iter_batches(columns=["text"], batch_size=2000):
+            for x in b.column("text"):
+                s = x.as_py()
+                if s:
+                    yield s
+
+def write_shard(path, arr):
+    h = np.zeros(256, dtype=np.int32); h[0] = 20240520; h[1] = 1; h[2] = len(arr)
+    with open(path, "wb") as f:
+        f.write(h.tobytes()); f.write(arr.tobytes())
+
+def main():
+    tok = Tokenizer.from_file(TOKJSON)
+    EOT = tok.token_to_id("<|endoftext|>")
+    print(f"EOT={EOT} | shard={SHARD:,} | val 1/{VAL_EVERY} | chunk={CHUNK}", flush=True)
+
+    buf = np.empty(SHARD + 2_000_000, dtype=np.uint16); tn = 0; sidx = 0
+    val = []
+    di = 0; total = 0; t0 = time.time()
+    chunk = []
+
+    def flush_chunk():
+        nonlocal tn, sidx, di, total
+        if not chunk:
+            return
+        for e in tok.encode_batch(chunk):     # rayon -> all cores, single process
+            ids = e.ids
+            # EOT PREFIX (BOS before each doc) -> matches modded-nanogpt align_to_bos
+            if di % VAL_EVERY == 0:
+                val.append(EOT); val.extend(ids)
+            else:
+                buf[tn] = EOT; tn += 1
+                m = len(ids)
+                buf[tn:tn+m] = ids; tn += m
+                if tn >= SHARD:
+                    write_shard(f"{OUT}/polish_train_{sidx:06d}.bin", buf[:tn])
+                    total += tn; sidx += 1; tn = 0
+            di += 1
+        chunk.clear()
+        print(f"  {di:,} docs | {(total+tn)/1e9:.2f}B train tok | {len(val)/1e6:.1f}M val | {time.time()-t0:.0f}s", flush=True)
+
+    for s in doc_stream():
+        chunk.append(s)
+        if len(chunk) >= CHUNK:
+            flush_chunk()
+    flush_chunk()
+    if tn:
+        write_shard(f"{OUT}/polish_train_{sidx:06d}.bin", buf[:tn]); total += tn; sidx += 1
+    write_shard(f"{OUT}/polish_val_000000.bin", np.array(val, dtype=np.uint16))
+    print(f"\nDONE: {sidx} train shards, {total:,} train tok, {len(val):,} val tok | "
+          f"{di:,} docs | {time.time()-t0:.0f}s", flush=True)
+
+if __name__ == "__main__":
+    main()

src/train_bpe.py

@@ -0,0 +1,62 @@
+#!/usr/bin/env python3
+"""Train a Polish byte-level BPE (32k) on a domain-balanced sample of the corpus,
+then report fertility vs GPT-2 BPE. Runs where the parquet shards live (slayer)."""
+import glob, sys, time
+import pyarrow.parquet as pq
+from tokenizers import Tokenizer, models, trainers, pre_tokenizers, decoders
+
+DATA = "/home/ubuntu/dynaword/data"
+OUT = "/home/ubuntu/dynaword/polish_bpe_32k.json"
+CAP = 220 * 1024 * 1024   # ~220 MB text per source -> balances away the 71% legal skew
+VOCAB = 32768
+
+def files():
+    return sorted(glob.glob(f"{DATA}/*/*.parquet"))
+
+def balanced_texts():
+    for f in files():
+        src = f.split("/")[-2]; got = 0; done = False
+        for batch in pq.ParquetFile(f).iter_batches(columns=["text"], batch_size=1000):
+            for x in batch.column("text"):
+                s = x.as_py()
+                if not s:
+                    continue
+                yield s
+                got += len(s)
+                if got >= CAP:
+                    done = True; break
+            if done:
+                break
+        print(f"  sampled {src}: ~{got/1e6:.0f} MB", file=sys.stderr, flush=True)
+
+t0 = time.time()
+tok = Tokenizer(models.BPE(unk_token=None))
+tok.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=False)
+tok.decoder = decoders.ByteLevel()
+trainer = trainers.BpeTrainer(
+    vocab_size=VOCAB, min_frequency=2,
+    special_tokens=["<|endoftext|>"],
+    initial_alphabet=pre_tokenizers.ByteLevel.alphabet())
+print("training BPE...", file=sys.stderr, flush=True)
+tok.train_from_iterator(balanced_texts(), trainer=trainer)
+tok.save(OUT)
+print(f"trained in {time.time()-t0:.0f}s | vocab={tok.get_vocab_size()} | saved {OUT}")
+
+# fertility vs GPT-2 on a held-out-ish sample (later docs of wikipedia, general domain)
+import tiktoken
+gpt2 = tiktoken.get_encoding("gpt2")
+sample = []
+for b in pq.ParquetFile(f"{DATA}/wikipedia/wikipedia.parquet").iter_batches(columns=["text"], batch_size=1000):
+    for x in b.column("text"):
+        sample.append(x.as_py())
+    if len(sample) >= 6000:
+        break
+sample = sample[4000:6000]                     # avoid the head used in training
+words = sum(len(s.split()) for s in sample)
+chars = sum(len(s) for s in sample)
+ours = sum(len(e.ids) for e in tok.encode_batch(sample))
+g2 = sum(len(x) for x in gpt2.encode_ordinary_batch(sample))
+print(f"\nFertility on {len(sample)} held-out PL docs ({words:,} words):")
+print(f"  polish-32k : {ours/words:.3f} tok/word | {ours/chars:.3f} tok/char")
+print(f"  gpt2-50k   : {g2/words:.3f} tok/word | {g2/chars:.3f} tok/char")
+print(f"  -> {g2/ours:.2f}x fewer tokens with the Polish BPE")