Datasets:
govpl: drop 186 third-party texts by a pinned review table (v2#106)
#170
by ppuzio - opened
The govpl card said that 710 kept articles carry a "Źródło:" line naming some other source and were not reviewed. This PR reviews them and drops the third-party ones. Part of slayerlabs/polish-dynaword-v2#106.
What is dropped (186 rows, 645,423 characters, 240,785 tokens)
| reason in the review table | rows | characters |
|---|---|---|
credit_media: the text is credited, wholly or partly, to a news portal, news agency, broadcaster or magazine |
13 | 50,933 |
credit_other: the text is credited, wholly or partly, to a source other than a public office (NGO, international or EU body, company, university, museum, other website) |
167 | 525,799 |
third_party_text: interviews and reprints found by title and phrase rules; 5 name another outlet, 1 names none |
6 | 68,691 |
What stays (530 "Źródło:" rows)
| reason | rows | characters |
|---|---|---|
credit_public_office: the text is credited to a public office only |
436 | 1,139,378 |
no_text_credit: the line credits data, a chart, a reference, a quotation or a photo, not the text |
94 | 687,410 |
The 40 teasers that quote a third-party interview and point to it also stay.
How
data/govpl/govpl.third-party-review.csv:id,decision,reason, 716 rows.src/clean_govpl.pypins the table by SHA-256 and drops itsdroprows after the PAP rule, from the same pinned v0.2.5 Parquet. Ids of the other rows are unchanged.- The build stops if the table and the
Źródło:regex disagree on any row, or if a listed id is not in the input. - New stats:
drop_third_party,drop_third_party_by_reason,review_sha256. - Checked: the new shard and sidecar equal main's minus exactly the 186 ids.
Rows and figures
| rows | characters | tokens | |
|---|---|---|---|
| main | 88,002 | 217,515,240 | 80,445,758 |
| this PR | 87,816 | 216,869,817 | 80,204,973 |
The card's legal note, filters, statistics and attribution count (1,189 → 1,185 rows fetched from several URLs) are updated, together with the registry entry in src/sources.py and a note in artifacts/source_findings.md.
Limits
- The classification is by the source the line names (141 groups), not a reading of every article.
- Misses outside the rules: 0 in a seeded sample of 200; the 95% upper bound is 1.83%. No larger read was done.
Tests
src/test_govpl_contract.pyadds a regex test, a test that the shard'sŹródło:rows are exactly the table'skeeprows and that nodropid is left, and a check that the card carries the registry text. 15 passed.pytest srchas no new failures compared with main. The govpl shard tests that error in a checkout without LFS objects now run and pass.
Dropping rows is a content change; it is decided at the merge.
🤖 Generated with Claude Code
ppuzio changed pull request status to open
ppuzio changed pull request status to merged
ppuzio deleted the
refs/pr/170 ref