Alexey5676/russian-supreme-court-plenum-acts

Plenum Resolutions of the Supreme Court of Russia (1961–2026)

Every act published in the «Постановления Пленума» section of the Russian Supreme Court's website: 1,504 records — 1,503 plenum resolutions plus 1 meeting agenda — with full texts, metadata and the court's original attachments. Coverage 1961–2026; completeness verified against the court's own index at collection time (the section reported exactly 1,504 documents).

Постановления Пленума ВС РФ — руководящие разъяснения по вопросам судебной практики, обязательные для всех судов. На момент сбора открытого машиночитаемого корпуса этих актов не существовало ни на Kaggle, ни на Hugging Face, ни на GitHub.

Files

Notebook

`notebooks/rag_plenum_vs_rf.ipynb` - building a retrieval system over this corpus, curation first. It derives from the corpus itself which acts are actually judicial guidance (572 of the 1,504 are bills submitted to the State Duma, 268 are organisational, 25 are compliance reviews; 363 are guidance), which ones the Court has since repealed (from its own stamps and from repeals declared inside later acts), which ones are still formally in force but interpret abolished codes (УК РСФСР, КЗоТ — dead in substance, with no stamp to say so), and which individual points a later act amended - then chunks the 215 acts that survive all three checks by their own numbered points, tags them on three deterministic axes (branch of law from code citations, legal institute from a closed vocabulary, and the article of the code a point relies on — bound to the right code, so «ст. 333 ГК РФ» never collides with «ст. 333 АПК РФ»), indexes them (BM25 over stemmed Russian + multilingual-e5, fused with RRF) and measures three ways of selecting the corpus on one labelled question set.

A second, independent benchmark measures what the first one hides - whether the right point is found, not just the right act. It is mined from the Supreme Court's own practice reviews, where a legal proposition is stated and the governing point is cited right after it, and is published here as benchmarks/plenum_points_from_reviews.json (297 pairs, 236 points, 76 acts, 73 source reviews, CC0 like the corpus). On it the system puts the right act first about half the time and the right point first about a quarter of the time. Because it is roughly seven times wider than the point-level set used previously (66 acts against 9), it also overturns three earlier conclusions of this notebook: point-level R@1 is 0.24 rather than 0.10, the hybrid is no longer beaten by plain BM25 on points, and two-stage retrieval (pick the act, then rank its points) loses depth rather than gaining it - its first stage drops the right act in 40% of queries. Retrieved points carry an "amended by act N" flag rather than being silently rewritten, and repealed acts are flagged and filtered rather than deleted. It opens with a sample result and closes with a runnable ask() sandbox. Text in Russian.

Field dictionary

Collection method

Collected 2026-09-13 from www.vsrf.ru with a polite crawler (~1.6 s/request): the site's own document-search endpoint (the one its frontend uses) for the index, static /upload/ files for attachments. Texts: 957 from the PDF text layer the court embeds (including digitized Soviet-era scans), 547 scraped from document detail pages — older records carry no attachments at all.

Text extraction

text is normalized, not verbatim: PDF extraction leaves page-break characters and pads lines with runs of spaces to justify them. Runs of horizontal whitespace are collapsed to one space, page breaks become blank lines, other control characters are dropped. Line breaks are kept as extracted, and hyphenated line-end words are left alone — undoing them would fuse genuinely hyphenated compounds. The court's original files are in the attachments archive for anyone who needs the raw layer.

Known limitations

• —53 records carry only their heading as text (text_completeness = title_only, mostly 1961–1986). Their document pages on vsrf.ru hold no body text and they have no attachments, so the full text is not public at the source. Filter them out for text tasks.
• —`category` is a heuristic over title and annotation, not an official taxonomy: the court files substantive rulings and organizational acts (bills submitted to the State Duma, court compositions, appointments) under one section. Expect edge cases.
• —`date_accepted` is empty for 1,157 records — the court's index omits it for older entries; date_published is always present.
• —Pre-2016 texts come mostly from detail pages; PDF-derived texts preserve the court's own OCR.

Legal status

Judicial acts are official documents and are not protected by copyright: subpara. 1 of para. 6 of Art. 1259 of the Civil Code of the Russian Federation excludes «судебные решения, иные материалы законодательного, административного и судебного характера» from copyright protection. The court publishes them itself under Federal Law № 262-FZ of 22.12.2008 on access to information about the activities of courts. Released under CC0 1.0. Source: Верховный Суд Российской Федерации, www.vsrf.ru.

Texts name judges and officials acting in their official capacity (signatories such as the Chief Justice and the Secretary of the Plenum); they carry no private-party personal data of the kind found in ordinary case files.

Provenance

Snapshot of 2026-09-13. This is a point-in-time export, not a live mirror: re-collection is manual, and each revision's date is stated here. The Plenum section changes slowly — new resolutions appear a few times a year.