# studia_bas Peer-reviewed Polish articles from *Studia BAS*, the socioeconomic and public-policy journal of the Bureau of Research (Biuro Analiz Sejmowych) at the Chancellery of the Sejm. Each record is one article, extracted from the text layer of the official issue PDF. ## Dataset description - **Original platform:** https://studiabas.sejm.gov.pl/journal.nsf/pubBAS.xsp?view=12&seria=4&lang=PL - **Publisher:** Kancelaria Sejmu, Biuro Analiz Sejmowych - **Domain:** academic, economics, social policy and public policy - **Language:** Polish (`pl`) - **License:** `CC-BY-3.0-PL` — https://creativecommons.org/licenses/by/3.0/pl/ - **Created:** 2021-2025 (per article in the `created` column) - **Added:** 2026-07-19 (inherited from the global `ADDED` constant in `src/sources.py`, which is what `build_dynaword.py` stamps; the shard is left as the committed pipeline produces it rather than hand-restamped) ## Statistics | read | kept | characters | tokens (`cl100k_base` proxy) | issues | |---:|---:|---:|---:|---:| | 115 | 112 | 6,089,085 | 2,256,009 | 17 | The released Parquet uses the canonical DynaWord schema: `id, text, source, added, created, token_count, license, author` Every retained article carries its author list in `author`, read from the PDF title page. Because the canonical schema has no column for it, the per-article DOI, attribution URL and source-PDF hash that CC BY attribution requires live in `studia_bas.attribution.jsonl`, one row per Parquet row, in the same order. ## Provenance and Licensing - **Origin**: Official journal archive at https://studiabas.sejm.gov.pl/journal.nsf/pubBAS.xsp?view=12&seria=4&lang=PL - **License**: Creative Commons Attribution 3.0 Poland (CC BY 3.0 PL). - **License Evidence**: CC BY 3.0 PL applies from issue 2(66)/2021 onward. Issues 1(1) through 1(65) are "Wszelkie prawa zastrzeżone" and are strictly excluded. This is corroborated by the journal's own statement and by pl.wikipedia.org/wiki/Studia_BAS. Eligible range is issues 66-82 = 17 issues. All 17 source PDFs are present locally and their SHA-256 hashes and byte sizes were independently recomputed and matched 17/17 against `data/studia_bas/studia_bas.license-evidence.json`. - **Genre/Register**: Peer-reviewed socioeconomic and public-policy articles, academic register. - **Coverage**: COMPLETE for the eligible range. Cross-checked against Crossref (journal ISSN 2082-0658, 246 works). Crossref lists 132 DOIs for issues 66-82; the corpus has 115. The 17-record difference is fully explained: every one of those 17 DOIs has type 'journal-issue' (issue-level front matter, no title, no authors, no pages), not 'journal-article'. Zero journal-article records are missing, and zero corpus DOIs are absent from Crossref. - **Coverage Cutoff**: `2(82)/2025`. Crossref holds nothing above DOI 10.31268/StudiaBAS.2025.10; the whole 2025.11-2025.25 and 2026.x ranges return HTTP 404. Newest Crossref deposit is 2026-03-06, i.e. roughly a 9-month deposit lag, so later issues may exist without DOIs yet. ## Collection and Filtering - Articles are extracted from the text layer of the official issue PDFs. They are split by DOI title pages. - **Normalization/Modification**: - Running headers, page numbers, author affiliation/contact blocks, terminal bibliographies, and ORCID values are excluded. - Common PII (emails, national identifiers) is replaced with `[PII]` to retain sentence structure. - PDF line wrapping is normalized without flattening paragraph breaks. - **Drops (Language Gate)**: 115 read -> 112 kept. The three language-gate drops are genuine English-language articles, correctly dropped because they contain insufficient Polish: - DOI 10.31268/StudiaBAS.2021.26 ("The EU Fiscal Risk Matrix") - DOI 10.31268/StudiaBAS.2022.08 ("Cultural and creative sectors in the COVID-19 pandemic crisis") - DOI 10.31268/StudiaBAS.2025.08 ("Integration of electricity markets throughout the European Economic Area") ## Argument for Inclusion This source adds high-quality, professional, and peer-reviewed academic texts in the domain of economics, social policy, and public policy, enriching the corpus with formal legal and socioeconomic Polish terminology. ## Metadata Quality - **Record ids** are derived from the article DOI (`studia_bas__`) and are stable across rebuilds (`preserve_id`). - **Authors** are read from the PDF title page, not from any metadata service. All 112 released author lists were cross-checked against Crossref: 108 match exactly, and none of the 4 differences is an extraction error. Three are errors in Crossref itself, where the PDF is correct ("Dorora"/"Dorota" Wyszkowska in 10.31268/StudiaBAS.2021.36, "Maciej N."/"Maciej J." Nowak in 10.31268/StudiaBAS.2023.02, and a dropped middle initial in 10.31268/StudiaBAS.2024.02); the fourth, 10.31268/StudiaBAS.2021.23, is a Crossref record that carries no author field at all. Co-authors are delimited by the affiliation markers the journal prints on the title page rather than by line spacing, because the journal sets a wrapped name and a following co-author at the same pitch; `src/test_studia_bas_contract.py` pins both that rule and the case of an author list running past the ordinary title-page band. - **PII audit**: `src/audit_studia_bas.py` finds zero emails, ORCIDs, PESEL, NIP or REGON values across all 115 extracted records. Run `src/report_studia_bas_audit.py --corpus ` to reproduce it. The report also lists advisory extraction-quality heuristics; those matchers are deliberately broad and fire on ordinary inline references, so they do not fail the audit. ## PII scrubber and NERGAL gate Scrubber that ran: `scrub_pii()` in `src/fetch_studia_bas.py` — regex pass over the extracted text (e-mail, ORCID, labelled PESEL/NIP/REGON, labelled phone -> `[PII]`), plus structural exclusion of author-affiliation/contact blocks and ORCID values at extraction time. This is **not** [SlayerLab/NERGAL](https://huggingface.co/SlayerLab/NERGAL); per AGENTS.md the shard must be re-run through NERGAL before it is admitted to a release (quoted case material and third-party names can survive any regex pass). Merge with `release: None` is fine meanwhile. The independent audit above (0 structured residuals over all 115 records) is evidence for the current build, not a substitute for the model-based pass.