ERrag: SAG benchmark conversion

This repository contains a JSON conversion of the public EnterpriseRAG-Bench test split for the SAG benchmark. The source dataset describes a simulated company; it is not a dump of real enterprise records.

Files

• errag_corpus.json: 511,961 documents, each with doc_id, title, and text.
• errag.json: 500 questions, each with id, question, answer, and expected_doc_ids.
• id_remap.json: four original document IDs mapped to stable -2 IDs for their second, distinct occurrence.
• excluded_documents.json: one source row omitted because its content is empty. It is not referenced by a question gold ID.
• sample_errag_corpus.json: 28 sample documents in the same converted doc_id / title / text format.
• sample_errag.json: 10 sample questions in the same evaluation format.
• errag_mid10k_corpus.json: 10,000 documents in the same converted format: the 722 IDs referenced by the question gold lists (723 gold references, including both occurrences of the duplicated ID used by qst_0413), plus 9,277 further documents drawn from the remaining errag_corpus.json documents. The selection uses a fixed seed (20261005), so re-running the conversion reproduces the same corpus.
• errag_mid10k.json: the same 500 questions as errag.json, unchanged, so the question-type distribution and gold lists are identical. It is meant for runs where the full 511,961-document corpus is too slow but the 28- document sample is too small.

Source and sample files

The datasets/ directory contains the source and sample dataset files:

• datasets/EnterpriseRAG-Bench/: the original benchmark README, inspection summary, and test-split Parquet files (511,962 document rows; 500 questions).
• datasets/sample_errag/: sample test-split Parquet files (28 documents; 10 questions).

The files under datasets/EnterpriseRAG-Bench/ retain the source dataset's original IDs. The root JSON files are the SAG conversion described below.

The source has 511,962 document rows. Both distinct documents are preserved for each duplicated doc_id: the first keeps its original ID and the second receives an -2 suffix. If a question lists that ID twice, its two gold references become [original ID, original ID-2] in source order. In particular, question qst_0413 references both versions. Questions with no document gold IDs are retained; 30 of the 500 questions have an empty expected_doc_ids list, so document recall is undefined for those questions.

Document content is otherwise retained as in the source Parquet, including any serialized-list text. The conversion script and validation tests are in the SAG benchmark project; this repository holds the resulting data.