jinulee-v/expert-rag-benchmarks

Expert RAG Benchmarks

A unified collection of four expert-level legal RAG benchmarks, exposed as six named splits and three relational configurations: questions, documents, and qrels.

The KCL split is named kcl_essay because Hugging Face split identifiers do not permit hyphens; its source name remains kcl-essay.

Loading

from datasets import load_dataset

repo_id = "jinulee-v/expert-rag-benchmarks"
questions = load_dataset(repo_id, "questions", split="housing")
documents = load_dataset(repo_id, "documents", split="housing")
qrels = load_dataset(repo_id, "qrels", split="housing")

Join questions.qid to qrels.qid, then join qrels.doc_id to documents.doc_id.

Sizes

Split	Questions	Documents	Qrels
housing	6,853	1,837,403	17,862
barexam	1,195	856,835	1,195
kcl_essay	169	128,790	531
legit_train	24,262	128,789	87,423
legit_val	200	128,789	721
legit_test	300	128,789	1,229

The kcl-essay, legit_train, legit_val, and legit_test splits use the shared 128,789-document Korean legal corpus. KCL additionally includes the 2020모633 judgment crawled from the National Law Information Center, bringing its corpus to 128,790 documents. Each corpus is exposed under its applicable split to keep the loading interface uniform.

Schemas

• questions: qid, question, answer, answer_type, choices_json, eval_json, meta_json, gold_contexts_json
• documents: doc_id, title, text, meta_json
• qrels: qid, doc_id, relevance, citation, issue_id

Fields ending in _json contain JSON-encoded values. Encoding heterogeneous nested metadata this way preserves it losslessly while maintaining one stable Parquet schema across all six splits. Decode one with json.loads(value) when the value is not null. citation is populated by Housing QA; issue_id is populated by LEGIT; both are nullable elsewhere.

Benchmark notes

• housing: yes/no questions about U.S. state housing law as of 2021. The full retrieval corpus contains statutes from 48 states.
• barexam: multiple-choice MBE questions over the complete annotated passage pool. The upstream train/validation/test examples are combined here as the requested barexam split.
• kcl_essay: Korean bar-exam essay questions with instance-level rubrics. A reviewed correction spreadsheet adds 13 unique relevance labels (531 total, covering 164 of 169 questions). Reference judgments are used where the fixed ID column was blank; 2020모633 was crawled from law.go.kr. The missing Constitutional Court judgment 2013헌바271 remains excluded.
• legit_train, legit_val, legit_test: Korean judgment prediction with legal issue-tree rubrics.

The metadata recorded with each source split is available under metadata/. The published legit_val metadata records 721 qrels, matching the file; the local precursor metadata had a stale count of 702.

legit_train contains 24,262 question rows but 24,086 distinct qid values: 176 identifiers each occur twice, with different question records. These source rows are preserved rather than silently deduplicated. Consumers that require a unique row key should combine qid with the row index; qrels continue to join on the source qid.

Sources, licenses, and attribution

This repository is a collection. The applicable license depends on the split:

• Housing QA and Bar Exam QA: CC BY-SA 4.0, from Stanford RegLab.
• KCL: CC BY-NC 4.0, from Korean Canonical Legal Benchmark (arXiv:2512.24572).
• LEGIT: CC BY-NC 4.0, from LEGIT (arXiv:2512.01020).

Users are responsible for following the license that applies to each split. The non-commercial restriction applies to KCL and LEGIT. The Housing QA and Bar Exam QA share-alike requirements continue to apply to those materials.

Intended use and limitations

The collection is intended for research on retrieval and retrieval-augmented generation in expert legal domains. It is not legal advice. Housing questions and statutes describe law as of 2021 and may not reflect current law. Benchmark answers, annotations, and retrieved materials can be incomplete or erroneous.