justatom/legal-benchmark

Legal retrieval: public working snapshot, 10,000 questions

Usable for pilot training and development; not a final gold benchmark or a hidden test. This snapshot contains 5,000 English and 5,000 independently generated Russian queries for retrieving English CJEU judgment paragraphs: EN→EN and RU→EN. It is not a Russian-law corpus. All splits contain equal numbers of EN and RU queries.

Split	Cumulative queries	New batch only
Train	7,526	836
Validation	1,020	0
Development-test	1,454	0
Total	10,000	836

The new batch is a subset of the cumulative dataset; do not concatenate both. There are 5,000 cumulative bilingual families and 418 new families. The corpus contains 593,877 English paragraphs and is included once in corpus/0000.parquet. docno is the corpus ID; text is the paragraph text.

Loading

The repository is public; no Hugging Face login or access token is required to read it.

from datasets import load_dataset

repo = "justatom/legal-benchmark"
queries = load_dataset(repo, "queries")
train_en = load_dataset(repo, "train_en", split="train")  # 3,763 rows
train_ru = load_dataset(repo, "train_ru", split="train")  # 3,763 rows
qrels = load_dataset(repo, "qrels")
corpus = load_dataset(repo, "corpus", split="corpus")

For reproducible experiments add revision="working-10000-v1" or the verified commit SHA to these calls. The earlier working-1164-v1, working-2164-v1, working-3164-v1, working-4164-v1, working-5164-v1, working-6164-v1, working-7164-v1, working-8164-v1 and working-9164-v1 tags are preserved. The new_queries configuration contains only the latest 836 questions, all in train. Typed empty validation/development Parquet files are retained but are not advertised as splits of this increment. The HF split development_test corresponds to release_split="development-test" in the rows. It has already been used during development and must not be described as an untouched or hidden test.

Raw JSONL/TSV snapshots are available under snapshots/cumulative-10000/ and snapshots/new-836/. The earlier cumulative-1164 through cumulative-9164 snapshots are retained as archives. snapshots/new-1000/ now archives the preceding batch; it is not the latest increment. For justatom, use train_pairs.en.jsonl or train_pairs.ru.jsonl with content_field=positive, labels_field=query, chunk_id_col=positive_id. Packaged JSON configuration paths are relative to the downloaded repository root; run from that root or resolve paths explicitly.

Label semantics and training precautions

• Each query has one model-audited anchor positive, not an exhaustive qrel set. Other corpus paragraphs are unjudged, not verified negatives.
• EN/RU siblings share an anchor. Do not treat them as negatives in diagonal InfoNCE; use separate-language batches or appropriate multi-positive masking.
• training_candidate_docnos.jsonl excludes globally reserved validation/test case IDs. These are case-safe candidates, not judged negatives. Mine on train only and filter likely relevant documents.
• Query-source and anchor case IDs do not cross split boundaries. Semantic overlap across splits is observed, not merely a theoretical possibility. For example, train LB-3b3bef51d006f0f79725-RU, validation LB-261d3d129b7904aebe2a-RU, and development-test LB-7191acf1aa9510955978-RU paraphrase the same question about interpreting EU law using its wording, context and objectives. Current exact/shingle checks do not eliminate equivalent legal questions. This is not a semantically independent test; cluster-level re-splitting and alternative relevance judgments are required for a separate rigorous research release.
• Questions were generated and audited using separate calls to GPT-5.6 Sol, with deterministic checks and targeted revisions. This is not independent legal-expert review. Some generic-regime wording remains weak or ambiguous.
• The new batch included assistant inspection of 120 candidate questions; 67 of those were retained. This includes 20 preselected questions, targeted difficult cases and all 68 replacement questions. It is not a random estimate of population error rate or proof of all-row quality.
• Some indexed anchor paragraphs depend on surrounding judgment context. An anchor can be relevant without being a self-contained answer paragraph. Descriptive regime wording and EN/RU legal terminology still need expert review.

The dataset includes factual and technical questions arising in legal cases, not only questions about legal rules. For example, retained family LB-42da7bae08cf93e1a4e7 concerns acetyl content and modification of starch; the citing judgment uses those facts in a customs-classification context.

Some audit/revision feedback was contradictory: required legal subject wording was later rejected as leakage, or a proposed clarification was subsequently rejected as external knowledge. Those cases and costs were preserved; affected families were excluded where recorded assistant review found a material problem. Passing software tests or a model audit does not resolve these labeling limitations.

Verification and provenance

Annotation pipeline revision: 57682b5. The new 836-question batch and the cumulative snapshot passed local row/qrel checks, case-overlap checks, global question-duplicate checks, recorded-response replay and actual justatom loader checks. Of 452 completed candidate families, 34 were excluded: three had a language in model quarantine and 33 had assistant exclusions; these groups overlap in two families. Four completed replacement rounds generated 20, 9, 3 and 2 bilingual families. The final 418 retained families contribute 836 questions, all assigned to train under the frozen case partition. No model-passed row was silently rewritten to meet the target. No normalized question duplicate remained against the prior published questions. All 5,000 retained positive paragraph texts were checked against the actual corpus; the previous 9,164 exported question records remain unchanged.

The current batch used 150 completed model calls, recorded at $20.1028512 in API-equivalent annotation cost (3,253,807 input tokens including 1,071,488 cached, 547,249 output tokens). This is not a billing statement and excludes this chat's manual review/orchestration usage and earlier batches. There were no new known annotation timeouts with missing usage in this batch. Recorded usage includes all generation, audit and revision calls in the primary and replacement runs.

The extended source pool omitted one export metadata field. An export-only, versioned exact-ID join restored source_audit_semantic_fit from saved audits; frozen generation inputs, queries, corpus and old exported records were not edited. Original source and derived-file hashes are recorded separately. This restoration required no new annotation calls.

reports/verification-summary.json records the pre-upload checks. reports/new-836-evaluation/ contains BM25 and Qwen3-Embedding-0.6B full-corpus smoke rankings and metrics for the new 836 questions only. These are anchor HR@1/2/5 and full-rank anchor MRR, not exhaustive relevance metrics and not an AnchorBank-vs-InfoNCE comparison.

upload_manifest.json records file sizes/checksums and source snapshot hashes. This package is a versioned working snapshot; it does not claim to satisfy the separate final-release/legal-expert-review gates of the source project. Raw model event logs, credentials, local machine paths, quarantined questions, and future-split source family inventories are not part of this upload.

Attribution and license

Adapted from LegalPincite, by Theresia Veronika Rampisela, Henrik Palmer Olsen and Giovanni Colavizza; DOI 10.57967/hf/9072. The corpus is the unchanged doc_par/test/0000.parquet shard at converted revision 7344efcd381a1083af3736fbcc585824ceb781c2; SHA-256 55a2d5e374b5203ce8b9eabc22e4508bea10236194cd446728d459c67830a949.

Data and adaptations are distributed under CC BY 4.0. Changes include selected anchor pairs, synthetic EN/RU questions, benchmark IDs, splits, model audits and storage formats. See DATA_NOTICE.md and provenance.json for source revisions, required citations and limitations. No endorsement by the authors or EU institutions is implied. Russian queries are not official translations. This research data is not legal advice.