Datasets:
Dedupe labels and remove Structural rows
Removes rows from the labels config whose function_majority is Structural, and deduplicates rows.
- Input: 3,073,770 rows
- Exact duplicate rows: 0 (dedup was a no-op;
(document_id, section_index)was already unique) - Removed
function_majority == "Structural": 352,325 rows - Output: 2,721,445 rows. Schema and zstd compression are unchanged. Rows with a null
function_majority(16,372) are kept.
README: updated the labels row count, and the note that sections and labels share row order (they no longer align by position, so join on (document_id, section_index)).
The sections config is unchanged.
Update: also dedupes within each document by exact content match (joined from sections), keeping the first occurrence by section_index. Applied after the Structural filter.
- After Structural filter: 2,721,445 rows
- Removed within-document duplicate content: 14,355 rows
- Final: 2,707,090 rows
Duplicates across different documents (~40k extra copies, mostly shared model-code text across jurisdictions) are intentionally kept. The sections config is unchanged.
Update: the same filtering now applies to the sections table (the text table), so sections and labels stay aligned row-for-row.
- Drop rows with
function_majority == "Structural": 3,073,770 β 2,721,445 - Within each document, drop rows whose
contentexactly repeats an earlier section (keep lowestsection_index): β 2,707,090
Verified from the actual content: 0 Structural rows, 0 within-document duplicate content, all 2,360 documents still present, schema unchanged, and keys identical in order to labels. Cross-document duplicate text is kept. manifest.json is left as the original build record.