Datasets:
sejm_committee_transcripts: card says fourteen exact labels, not eleven
Card-only fix for sejm_committee_transcripts, follow-up to Hub PR 106 (ac7bd9f). No data change.
What was wrong: the card gave two different counts for the labels kept by exact match. The speaker-rules paragraph (line 114) said "Fourteen labels", and the cut paragraph (line 158) said "eleven labels with a typo or without a role word". EXACT_LABELS in src/fetch_sejm_committee_transcripts.py (line 319) has 14 entries. "Eleven" was the count before the second round of PR 106.
Change (1 file, +4/−3):
- Line 114: the 14 entries are of three kinds, not two. The two labels of the MON subcommittee chair (an MP) carry no "poseł", so the sentence now reads "with a typo in the transcript, without a role word, or without "poseł"". The counts (41 turns, 70,395 characters, 63,208 of them those two labels) are unchanged and were re-counted in the merged parquet.
- Line 158: now says "the fourteen
EXACT_LABELSabove", with no second count to drift.
Not changed: artifacts/.../source_findings.md line 2350 ("eleven … fourteen after the second round") is historically correct. The "typo labels" wording in the fetch.py and clean.py docstrings is code and out of scope.
Render sample: not applicable, because the parquet, attribution, stats and roles CSV are untouched.
Tests: src/test_sejm_committee_transcripts_contract.py passes 59 of 59 with the parquet and attribution.jsonl materialised from LFS (sha256 8c7d038e…, the artifact on main). The full suite in a skip-smudge worktree has the same failures as main, all of them LFS pointers.
🤖 Generated with Claude Code
refs/pr/107 ref