RuTaR — Russian Tax Reasoning

This is a format mirror of rutar-anonymous/RuTaR, attributed to its upstream authors. It contains Russian tax questions derived from letters of the Russian Ministry of Finance and Federal Tax Service. Original binary labels are preserved: 1 = yes, 0 = no.

Original source and attribution

• Upstream repository and documentation
• Exact upstream commit
• Original workbook
• Original reference-source corpus
• Individual government-letter links are retained in source_url, with document titles, dates and issuing body in each record. One source URL is missing upstream.

The upstream snapshot does not include a license file or a license statement. This mirror does not assert or grant a new license; consult the original repository/authors for reuse terms. No author identity, paper citation or license has been inferred from the anonymous repository name.

Files and preservation

• data/test.parquet: all 209 nonempty workbook records, in original order.
• original/rutar.xlsx: byte-for-byte original workbook (SHA-256 5ce0824ea3d806c2475ac2d1880cfd1795a34ff1ac31a65b94a14d75cd136dc7).
• sources/sources_dataset_for_rutar.json: byte-for-byte original RAG corpus, 480 citation-to-text entries.
• manifest.json: upstream revision, file checksums and conversion version.

test is a packaging split for the complete upstream table; upstream supplies no train/test split. No questions, duplicates or gold labels are removed here. Blank question at ID 119 and duplicate questions at IDs 114/117 remain available for auditing. Empty worksheet rows are omitted.

The unnamed workbook index becomes a stable string id. true_answer becomes an integer 0/1. Other original columns remain strings with blank cells represented as null. found_sources retains its original string representation of a list. Dates remain original strings. Original XLSX is retained for exact source recovery.

Columns

Column	Meaning
id	Original workbook numeric index, represented as a string
title	Government document title
date_publication	Original document publication date
full_text	Full original letter, including its answer
question_letter / answer_letter	Question/answer parsed from the letter
letter_type	Issuing body (minfin or Federal Tax Service identifier)
source_url	Link to the original government document
question_for_llm	Upstream prepared yes/no question (one null)
found_sources	Original string list of references; texts are in the source corpus
true_answer	Gold: 1 yes, 0 no

Loading

from datasets import load_dataset
rows = load_dataset("RefalMachine/RuTaR", split="test")

Pin revision to a Hub commit for reproducible evaluation. HF caches the Parquet snapshot locally, so ordinary repeated runs reuse downloaded files.

Evaluation considerations

The full letters and their parsed answers contain evidence used to author the questions; feeding them into a closed-book prompt leaks gold information. For closed-book evaluation, use only question_for_llm plus an explicit answer instruction, and keep demonstration questions/documents separate from evaluation. LLMTF reserves IDs 0–4 as a fixed demonstration pool, drops missing question 119 and duplicate 117, and evaluates 202 questions for both zero-shot and five-shot. That is an LLMTF protocol, not an upstream split or a reproduction of upstream RAG.

Labels refer to the source documents and their historical dates; they have not been rewritten to reflect current tax law. This mirror preserves upstream annotations and does not add legal expert validation.