feat(data): add reviewed Global Voices Polish corpus

## Summary

This PR adds `global_voices`, a directly acquired and license-reviewed source of Polish-language civic-media and public-interest journalism, to Polish DynaWord.

The release contains 2,040 articles published by the Polish edition of Global Voices between 2008 and 2023. Global Voices describes its Lingua translations as human-made work performed by a volunteer translator community. The source therefore contributes edited, translated journalism rather than conversational Polish or text necessarily authored originally in Polish.

## Dataset

- artifact: `data/global_voices/global_voices.parquet`
- documents: 2,040
- characters: 10,073,628
- tokens: 3,715,020 (`cl100k_base` proxy)
- date coverage: 2008-11-04 — 2023-05-24
- license: `CC-BY-3.0`
- schema: `id, text, source, added, created, token_count, license, author`
- Parquet SHA-256: `1ca02fdb6bad32b20757ddcd9b9f072056ff35b68c487143138cc7912b3a45c3`

## Provenance and cleaning

Articles were acquired directly from `https://pl.globalvoices.org/`, not from a third-party corpus. The release received a reviewed cleaning pass: redundant source-language passages were removed where the corresponding Polish translation was retained, together with extraction artifacts and obsolete interface text. Intentional quotations and source-language material required for context were preserved.

The source release contained 2,044 records. Four records below the minimum quality and length requirements were excluded, leaving 2,040 documents with no exact duplicates within the source. The pipeline introduced no machine translation or automatic speech transcription.

## Licensing and attribution

Each released article carries `CC-BY-3.0` metadata. The DynaWord `author` column distinguishes original authors from Polish translators where contributor information is available.

Because the fixed DynaWord schema does not contain article titles or URLs, `global_voices.attribution.jsonl` maps every released ID to its title, canonical article URL, contributor attribution, license URL and modification notice. The sidecar covers all 2,040 documents and has SHA-256:

`ab8ad393c4453524c4f3d7fe89157bfca1d9f66f2a735815f3b1ebd405c8fa13`

Binary media are not redistributed. Rights in separately quoted works, captions or embedded social-media text may remain with their respective authors.

## Limitations

This source represents edited online journalism and human translation, not conversational Polish or the full distribution of contemporary Polish. Translationese, source-specific editorial conventions and intentional multilingual quotations may remain. Coverage ends in May 2023. Cross-source deduplication remains part of the complete DynaWord build.

## Included files

- `global_voices.parquet`
- `global_voices.stats.json`
- `global_voices.md`
- `global_voices.attribution.jsonl`
- `NOTICE.md`

data/global_voices/NOTICE.md

@@ -0,0 +1,11 @@
+# NOTICE — Global Voices source
+
+The exported article text is attributed to Global Voices and the credited
+original authors and translators under Creative Commons Attribution 3.0.
+
+Source: https://pl.globalvoices.org/
+License: https://creativecommons.org/licenses/by/3.0/
+
+Per-record contributor attribution is stored in the `author` column. The accompanying `global_voices.attribution.jsonl` sidecar additionally preserves each released article's title, canonical URL, license URL and modification notice. Twelve historical records had no machine-readable contributor field; full-text review recovered Didem Tali and Khalid Ibrahim for two of them. Ten records without an identifiable contributor use the explicitly organizational attribution `Global Voices`; this must not be interpreted as a known individual author.
+
+Binary media files are not included. Articles may nevertheless contain captions, URLs, social-media posts and quotations from third-party works; rights in those materials may remain with their respective authors. Article-level CC BY 3.0 is not asserted independently for photographs, videos, embeds or separately quoted works.

data/global_voices/global_voices.md

@@ -0,0 +1,65 @@
+# global_voices
+
+Polish-language civic-media and public-interest journalism from [Global Voices](https://pl.globalvoices.org/), produced by its international community of writers and volunteer translators.
+
+## Dataset description
+
+- **Source (upstream):** https://pl.globalvoices.org/
+- **Domain:** civic media, news and public-interest journalism
+- **Language:** Polish (pl)
+- **License:** `CC-BY-3.0`
+- **Created (range):** 2008-11-04, 2023-05-24
+- **Added:** 2026-07-15
+
+Global Voices describes its Lingua translations as human-made work performed by a volunteer translator community. The dataset primarily represents edited, translated journalism rather than conversational Polish or text necessarily authored originally in Polish.
+
+## Statistics
+
+| documents | characters | tokens (`cl100k_base` proxy) |
+|---:|---:|---:|
+| 2,040 | 10,073,628 | 3,715,020 |
+
+The Parquet schema matches DynaWord:
+
+`id, text, source, added, created, token_count, license, author`
+
+Published files:
+
+- `global_voices.parquet` — cleaned article text and DynaWord metadata;
+- `global_voices.stats.json` — release statistics;
+- `global_voices.attribution.jsonl` — title, canonical article URL, contributor attribution, license URL and modification notice for every document;
+- `NOTICE.md` — licensing and attribution notice.
+
+## Provenance
+
+Articles were acquired directly from the Polish Global Voices website through its official publication interfaces, rather than from a third-party corpus. Each released ID is linked to the corresponding canonical article URL in `global_voices.attribution.jsonl`.
+
+The source release contained 2,044 records. Four records below the minimum quality and length requirements were excluded, leaving 2,040 documents. The final source contains no exact duplicate documents.
+
+## Cleaning
+
+The release received a reviewed cleaning pass. Redundant source-language passages were removed where the corresponding Polish translation was retained, together with extraction artifacts and obsolete interface text. Intentional quotations and source-language material required for meaning or documentary context were preserved.
+
+The cleaning process did not use machine translation or automatic speech transcription. It should not be interpreted as making every document strictly monolingual.
+
+## Licensing — traceable basis and attribution
+
+CC BY 3.0 was verified from the article-level license link published with each retained Global Voices article. Each released record carries `CC-BY-3.0` metadata. Contributor roles are preserved in the `author` column where available, distinguishing original authors from Polish translators.
+
+Because the fixed DynaWord schema does not contain article titles or URLs, the attribution sidecar maps every released `id` to its title, canonical URL, contributor attribution, license URL and modification notice. It contains one unique URL-bearing record for each of the 2,040 Parquet documents.
+
+Recommended attribution:
+
+> Global Voices; the original author(s) and Polish translator(s) listed in the attribution sidecar; canonical article URL; CC BY 3.0. Article text was extracted and cleaned; redundant source-language passages were removed where a Polish translation was retained.
+
+Binary photographs, videos and embedded media are not redistributed. Articles may contain quotations, captions or social-media text whose separate rights remain with their respective authors.
+
+## Limitations and recommended use
+
+- The dataset represents edited online journalism and human translation, not the full distribution of contemporary Polish.
+- Translationese and source-specific editorial conventions may be present.
+- Some intentional source-language quotations remain inside otherwise Polish articles.
+- Coverage ends in May 2023.
+- Cross-source and near-duplicate removal against other DynaWord sources remains part of the complete corpus build.
+
+Global Voices is best used as a complementary source of contemporary civic-media and journalistic Polish, rather than as a standalone representation of colloquial or natively authored Polish.

data/global_voices/global_voices.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ca02fdb6bad32b20757ddcd9b9f072056ff35b68c487143138cc7912b3a45c3
+size 5253559

data/global_voices/global_voices.stats.json

@@ -0,0 +1,15 @@
+{
+  "read": 2044,
+  "kept": 2040,
+  "drop_short": 4,
+  "drop_lang": 0,
+  "drop_dup": 0,
+  "drop_ocr": 0,
+  "chars": 10073628,
+  "tokens": 3715020,
+  "licenses": {
+    "CC-BY-3.0": 2040
+  },
+  "authors_with_value": 2040,
+  "stats_recomputed_from_parquet": true
+}