# OPEN ICM Polish publications pilot This target-specific slice is derived from the deterministic 100-record OPEN ICM pilot. - Acquired: 100; retained after source QA: 95 - Exact title/DOI overlaps with `biblioteka_nauki` excluded: 10 - Dropped in the licence review (v2#111; no licence basis beyond the repository's own field, or the year rule): 74 - Proposed records: 11 - Proposed tokens: 441,475 (`cl100k_base` proxy) - Characters: 1,203,239; author coverage: 100.0% - **License:** per-record CC BY or CC BY-SA variant, each with a licence basis (see "Licence basis") - Immutable source snapshot: https://huggingface.co/datasets/PiotrSty/open-icm-pl-publications/tree/c4a20022e6947f55f82c3354bb5c86609ba681b8 Per-record attribution, item URLs, license labels, normalized identifiers, bitstream metadata and checksums are preserved. The DynaWord proposal excludes all exact normalized-title or DOI matches found against 42,071 pinned `biblioteka_nauki` rows. Cross-source text deduplication and benchmark checks remain integration gates. Repository-generated TEXT can retain PDF extraction errors, omit figures and flatten tables. Item-level licensing does not establish the status of every quoted third-party passage. ## Origin and license at the origin - Origin portal: **OPEN ICM** (Interdisciplinary Centre for Modelling, University of Warsaw) - . Items are discovered through the official DSpace API (`/server/api/core/items`), each record pins its `landing_url`, e.g. . - License statement **on the portal** (not the mirror): every item carries `dc.rights` + `dc.rights.uri` in its origin metadata - e.g. item `0078ca93-45a5-4bef-912d-297c069ee6ab` (dropped in the licence review, see "Licence basis") declares `dc.rights: "Uznanie autorstwa 3.0 Polska"` with `dc.rights.uri: http://creativecommons.org/licenses/by/3.0/pl/`, readable without JavaScript at and rendered on the landing page. That field is set by the depositor. 7 of the 11 kept records additionally ship a per-item `license.txt` bitstream, pinned by checksum in `open_icm_pl.attribution.jsonl` (`license_evidence`); the files were not read (the bitstream endpoints serve a CAPTCHA page to anonymous clients). - Allowlisted labels only (fail-closed gate in the fetch script). The 11 kept records carry CC-BY-3.0-PL (1), CC-BY-SA-3.0 (1), CC-BY-4.0 (3) and CC-BY-SA-4.0 (6); the 85 records before the review carried CC-BY-3.0-PL (32), CC-BY-SA-3.0-PL (38), CC-BY-SA-3.0 (1), CC-BY-4.0 (7) and CC-BY-SA-4.0 (7). `Dozwolony użytek` and NC variants are rejected. ## Licence basis (v2#111) The repository label (`dc.rights`) is set by the depositor, so it is not a basis by itself. A record is kept only if a second source states the licence; `open_icm_pl.attribution.jsonl` carries `licence_basis`, `licence_basis_url` and `licence_basis_date` per record, and `open_icm_pl.licence-basis.json` carries the evidence per record and the review. | Basis | Records | What it is | |---|---:|---| | `file_note` | 2 | the text prints a CC licence with the label's elements (CC BY-SA 3.0 and CC BY-SA 4.0 printed in the text) | | `publisher` | 9 | a publisher page states a CC licence; the page was read on 2026-10-20 and its quote and SHA-256 are in `open_icm_pl.publisher-check.json` | The review (`open_icm_pl.review.csv`) dropped 74 of the 85 records: | Reason | Records | Characters | Tokens | |---|---:|---:|---:| | the text prints a CC licence with an NC or ND element, which the repository label does not have | 3 | 483,652 | 178,111 | | only the repository states the licence; the publisher page read states no CC licence, or an NC or ND one | 6 | 328,077 | 118,365 | | only the repository states the licence; no publisher statement for the article found | 56 | 3,833,763 | 1,484,031 | | only the repository states the licence; no publisher page could be read | 8 | 322,300 | 118,695 | | a second source states the licence, but the year the text was created is earlier than the first year of that licence version (the year rule, below) | 1 | 47,988 | 17,497 | Three kept records changed their `license` column (and nothing else) to the publisher's statement: `168531fb`, `5a22a282` and `6fe4e956` CC-BY-SA-3.0-PL to CC-BY-4.0. Every other column of every kept record is byte-identical to the previous release; the test recomputes this from the shipped Parquet. The year rule: a record that the other rules keep is dropped when the start year of its `created` is earlier than the first year of the licence version of the label it would carry after the relabel (CC 2.0: 2004; 3.0 and 3.0 PL: 2007; 4.0: 2013; the table is `licence_basis.RELEASE`, shared with `biblioteka_nauki`). It hit one record: `0078ca93`, created 2011, whose publisher page states CC BY-SA 4.0. It gets no relabel and is dropped. Not established: - The publisher statements are as read on 2026-10-20. For older articles a publisher may have applied a statement to its back issues later. - The year rule compares years only, and its table is a lower bound (a Polish port or a journal may have adopted a version later). `6fe4e956` was created in 2013 and its publisher page is dated 2013-06-24, before CC 4.0 was released in November 2013; it is kept with CC-BY-4.0. - Three statements are journal-level, not article-level (`5a22a282`, `635e0996`, `8c8e1344`). - For `8c8e1344` the publisher states CC BY without a version; the label stays the repository's CC-BY-3.0-PL. - The saved copies of the publisher pages are not committed (SHA-256 and size are). The PDFs of the three relabelled records were not downloaded. ## What the texts are - **Genre**: Polish open-access academic publications - 9 journal articles (pedagogy, law, economics, history, psychology, programming education) and 2 monographs (`3cc80540`, `5fc6e6dd`). One record per publication. - **Register**: formal academic written Polish with abstracts, bibliographies and byline blocks (authors, affiliations, ORCID, contact e-mails). Text is repository-generated from PDFs (`is_ocr`-like extraction): math/font glyph noise ("❬❜❡♥✐❛✳❛❧..."), hex-encoded UTF-16 fragments and flattened tables survive in places. - **Time span**: `issued` 2013 - 2024-09 (the 11 kept records; the deterministic 100-record pilot spanned 1998 - 2024-09, and 6,183 items are eligible in total - yields must not be extrapolated). ## Scrubber and residual-PII spot check Scrubber that ran: **pattern redaction in `src/fetch_open_icm_pl.py`** (e-mail -> `[REDACTED:EMAIL]`, labelled phone -> `[REDACTED:PHONE]`, ID/IBAN patterns; pipeline counts over the 85-record shard before the licence review: 112 e-mail and 4 labelled-phone redactions; the 11 kept records hold 15 `[REDACTED:EMAIL]` and 2 `[REDACTED:PHONE]` placeholders). This is **not** [SlayerLab/NERGAL](https://huggingface.co/SlayerLab/NERGAL); the shard must be re-run through NERGAL before it is admitted to a release. Merge with `release: None` is fine meanwhile. Residual-PII spot check (60 of 85 records, seed 20260928): 85 `[REDACTED:EMAIL]` placeholders confirm the redaction works on clean contacts; **one real residual contact was found** - a glyph-garbled work e-mail in `open_icm_pl_6ec903ca-db89-4a17-acc8-850420c468ac` ("JKolbusz @wsiz.rzeszow.pl", stray space before `@` in a line mangled by broken font encoding) that the pattern missed. Benign non-hits: 161 digit runs (DOIs, URL slugs), 5 26-digit runs (hex-encoded UTF-16 extraction garbage), 1 IPv4-like hit (a CiteSeerX DOI URL fragment), 28 ORCIDs and 25 institutional byline addresses. Residual risk is MODERATE - extraction noise breaks pattern matching (the missed contact is proof), which is precisely why NERGAL (model-based) is the release gate. The record with the missed contact (`6ec903ca`) was dropped in the licence review. A second scan of the kept records (an `@` or "(at)" outside the placeholders, unredacted phone-like and ID-like digit runs; made on the 12 records kept before the year rule dropped `0078ca93`, of which the 11 are a subset) found no personal contact; it found one publisher imprint phone number that the pattern missed (`5fc6e6dd`, "12 431-27-43", the office of Wydawnictwo Księgarnia Akademicka, next to a redacted phone and e-mail). It is a business number, not personal data, and is left in place, because the rows are unchanged except for the three `license` values. ## Normalization and reproducibility note `src/fetch_open_icm_pl.py` downloads the repository-generated TEXT bitstreams (MD5-verified against declared checksums) and applies Unicode/whitespace normalization, PII pattern redaction, language/quality filtering and exact/near deduplication. Raw->shipped before/after excerpts cannot be regenerated anonymously at review time: the OPEN ICM bitstream endpoints serve an interactive CAPTCHA page to anonymous clients (the acquisition recorded in `attribution.jsonl` went through it). The raw bitstreams are pinned by checksum per record; the extraction-artifact residue is visible in the corpus samples.