Sophia-bench

Sophia-bench is a patent citation-relevance retrieval benchmark. It contains 10,000 query publications, 75,050 candidate publications, and 66,469 unique directed query-document relevance pairs. All three configurations use the test split: these are relational tables, not three independently sampled evaluation sets. No training split is provided.

Loading

from datasets import load_dataset

queries = load_dataset("Questel-AI-lab/sophia-bench", "queries", split="test")
corpus = load_dataset("Questel-AI-lab/sophia-bench", "corpus", split="test")
qrels = load_dataset("Questel-AI-lab/sophia-bench", "qrels", split="test")

Use streaming=True for sequential access without downloading the whole configuration first. Long patent descriptions can make full materialization memory-intensive.

Source and provenance

The data comes from Questel's international patent database. Text is provided in English, either original or translated, as described in the accompanying paper. LLM-generated fields are additional text views, not examiner-authored text or gold annotations.

query_id and doc_id are dataset-specific publication identifiers in a shared namespace, such as sophia-000001. The same publication has the same identifier in both tables. These are not standardized patent publication numbers.

source_xpn preserves the original external lookup identifier. Formatting varies, kind codes can be absent, and values are not guaranteed unique. Do not automatically append a kind code such as A or A1.

Queries and corpus fields

Field	Type	Meaning
query_id / doc_id	string	Dataset-specific publication identifier, shared across queries and corpus.
source_xpn	string	Unmodified source publication lookup identifier.
publication_authority	string	Two-letter publication authority code, such as CN, US, EP, or WO.
publication_date	string	Source publication date, formatted as YYYY-MM-DD in UTC.
source_apl	string	Language label from the source metadata, most likely the original language of the priority filing (not necessarily the language of the text). Empty for some older documents.
ipc	list[string]	IPC classifications; may be empty.
title	string	Patent title.
abstract	string	Patent abstract.
claims	string	Patent claims.
description	string	Patent description.
adb	nullable list[string]	Text extracted using Questel's internal pattern-matching rules.
iclm	nullable list[string]	Source text view intended to select independent claims; may include dependent claims.
obj	nullable list[string]	Object-of-invention text extracted using Questel's internal pattern-matching rules.

Text, HTML, whitespace, list order, nulls, and empty values are preserved, not cleaned or replaced during export. WO denotes WIPO/PCT international publication and EP denotes the European Patent Office. Authority is not a statement about where a patent is granted or enforceable.

Additional query fields

Field	Type	Meaning
ai_title	string	LLM-generated title view.
ai_claim_summary	string	LLM-generated claim-summary view.
ai_obj	string	LLM-generated object-of-invention view.
ai_adv	string	LLM-generated advantages view.
ai_ab	string	LLM-generated abstract view.
ai_features	list[string]	LLM-generated feature strings.
ipc_section	nullable string	Unmodified source IPC section metadata.

LLM-generated views may contain inaccuracies.

Qrels

Each row is one unique ordered (query_id, doc_id) pair. All rows represent observed citation relevance; there are no explicitly judged negative pairs. An absent pair is unjudged, not established irrelevant.

Field	Type	Meaning
query_id	string	Foreign key into queries.
doc_id	string	Foreign key into corpus.
category	string	Mutually exclusive XY, A, or overlap.
source_relations	list[struct]	Observed {category, direction} entries, retaining original aggregate XY, fine-grained X/Y, and A memberships.

Directions are relative to the query: cited means the query cites the target; citing means the target cites the query. XY is the union of X and Y labels, not an independent legal category. An X or Y entry and its aggregate XY entry in source_relations are not separate citation events. Do not count list entries as independent citation events.

X, Y, and A are citation-category labels, not labels identifying whether the examiner or applicant supplied a citation.

Category overlap

For each query, let R_XY union its cited/citing XY lists and R_A union its cited/citing A lists. Exported categories are:

• XY: R_XY - R_A;
• A: R_A - R_XY;
• overlap: R_XY & R_A.

Exclusive category	Pairs
XY	30,260
A	33,653
overlap	2,556
Total	66,469

An overlap pair has both XY and A source labels and is stored once. The cause of the overlap cannot be resolved from the available annotations, which lack report- and claim-level context.

Self-matches and reciprocal pairs

267 query publications also appear in the corpus. A query's own corpus entry is never a relevant document, and it must be excluded from the candidate ranking at evaluation time, i.e. a query cannot retrieve itself.

Reciprocal retrieval pairs (A, B) and (B, A) are retained. Incoming and outgoing citation aggregation can produce both from one underlying citation. They are separate retrieval tasks, not necessarily mutual legal citations. Same-family relations are not filtered.

Evaluation protocol

Exclude the query's own corpus entry from its candidate ranking; it stays available as a candidate for other queries. All qrels rows are binary positives. For XY-only evaluation use XY + overlap; for A-only evaluation use A + overlap (the two sets intersect on overlap). Report which protocol you use.

The source retrieval setup joins cleaned title, abstract, claims, and description with newlines for the tacd view; list-valued views are joined before cleaning. Exported values remain raw so preprocessing must be declared by each evaluator. Bidirectional citation relevance does not enforce prior-art availability at the query date and is not a legal patentability determination.

Limitations and responsible use

• Citation labels are incomplete relevance proxies and lack claim/report context.
• Source publication numbering is not fully normalized.
• English translations and AI views may introduce factual or linguistic errors.
• Do not use overlap labels or unjudged pairs as automatically verified negatives.
• Future train/dev/test construction should prevent reciprocal-pair leakage and assess publication/family overlap using appropriately verified metadata.

License

The dataset license is CC BY-NC-SA 4.0: attribution, noncommercial use, and ShareAlike for adaptations.

Reference

Citation-Driven Multi-View Training for Patent Embeddings: QaECTER and Sophia-Bench. This release applies explicit annotation cleaning; do not assume that previously reported scores used the same self-match or overlap protocol.