Argentine Senate stenographic transcripts, 1998–2026: a speaker-attributed corpus

Licence, in short: CC BY 4.0 covers what this project adds — the structure, attributions, annotations and reference tables — not the words of the record, which are the Argentine Senate's and are not relicensed; the code is MIT. See LICENSE-DATA in the deposit. Every stenographic record (versión taquigráfica) the Argentine Senate serves for its sittings from February 1998 to September 2026, parsed from the chamber's own PDFs and, for most of 1998-2003, its HTML export into speaker-attributed passages. 818 sittings held and 817 parsed (the one that is not, an impeachment tribunal of December 1997, is a photocopy with no text in it), 440,445 rows, 244,624 passages of speech, 69,290 typed stenographer's notes, 25.28 million words of speech. Each passage carries the printed speaker label, the senator it resolves to, and the caucus that senator sat with on the day of the sitting. Coverage is complete from 2002 on; 1998-2001 is partial because the portal lists those sittings but no longer serves every file. The bundle holds the per-sitting Parquet tables, the resolved speakers, the parse record, the archived pages the early caucuses are rebuilt from, the reference tables (roster snapshots, hand-dated caucuses, observed authorities), the annotated accuracy sets (72 PDF pages and 24 stretches of the HTML era, annotated by a language model), the provenance manifest, the pipeline with its pinned environment, and the documentation. Accuracy figures, the checks behind them and the known limits are in the README inside the bundle. The source files are not redistributed: they are the Senate's to publish. raw_data_manifest.csv identifies each one by SHA-256 and source URL, and scripts/download.py fetches them again, so the corpus can be rebuilt from the sources. The words of the record, in the text column, and the Senate's own pages and lists shipped for checking are the Senate's and are not relicensed; LICENSE-DATA sets out exactly what each part covers. To cite a transcript, cite the chamber as publisher, with the sitting date and the per-sitting URL from the manifest.