Dataset
patent-legal-ir-graphrag
The Patent Legal IR GraphRAG release follows the Publicus retrieval format, providing dense corpus shards, BM25 documents with CID-keyed length and entry metadata, and BM25 postings arranged for inverted-index use.
Data preview
Dataset card
The patent-legal-ir-graphrag dataset is a Hugging Face release by publisher justicedao that packages a BM25 and vector retrieval index over patent and legal materials for information-retrieval research. It is distributed as Parquet shards alongside graph adjacency files intended for GraphRAG-style hybrid search.
What this dataset contains
The release combines keyword postings, vector chunks and graph adjacency files laid out for sparse BM25 retrieval over a corpus that includes patent text, CFR sections and MPEP references. Index metadata is provided under an `indexes/` directory, with matching postings stored as `data/bm25/postings/part-*.parquet` shards. The corpus size listed in the schema ranges between one thousand and ten thousand records, with a train split of 997,485 rows.
Structure and fields
Each shard row is described by string identifiers such as `cid`, `first_key`, `last_key`, `relative_path`, `schema_version` and `sha256`, together with numeric fields including `row_count`, `size_bytes`, `start_document_index`, `end_document_index`, `shard_id` and a `kind` label. BM25 layout parameters in the README list an `average_document_length` of about 1,771.78, a `max_query_terms` of 64, `posting_rows_per_record` of 4,096 and `terms_per_shard` of 4,096. Files are mirrored on IPFS, supporting partial downloads of only the BM25 shards or vector chunks that match a probe query.
How analysts use it
The accompanying scripts support BM25, hybrid and citation-aware re-ranking over topics such as 'duty of candor and good faith' and 'information disclosure statement 37 CFR 1.97', with explicit section ranking available via a flag. Reported checks include 9 of 9 rule-based end-to-end cases passing and 13 of 13 pytest integration tests passing. For investment workflows, the index is better suited to patent law and regulatory text retrieval than to commercial signal generation, so its main application is internal tooling for IP-aware due-diligence teams rather than a market-alpha dataset.
Documentation & references 1
Related datasets
Browse moreParsed eligibility criteria from 13,229 ClinicalTrials.gov studies representing the candidate pool gathered by two first-stage retrievers during TREC Clinical Trials 2021–2023, formatted as typed entity-relation graphs for reranking models.
The clinical-trials-patient-graphs dataset is a small tabular collection of patient-level clinical records distributed in Parquet format by the 2001jdev publisher. It organizes entities and relations extracted from trial data into structured rows spanning 2021 through 2023.
This dataset is a parsed snapshot of ClinicalTrials.gov records, distributed in Parquet format on the Hugging Face Hub under an unspecified license. It contains roughly 52,000 trial entries drawn from public registry filings.
clinical-trials-trec-qrels is a tabular relevance-judgment file distributed by the 2001jdev user on the Hugging Face Hub. It maps clinical-trial topic identifiers to NCT registry IDs with graded relevance scores.
The clinical-trials-trec-topics dataset is a small tabular collection of clinical-trial topic records distributed in Parquet format on the Hugging Face Hub. It appears to be derived from TREC clinical-trial retrieval topics, with rows partitioned by year.
SEC EDGAR 10-K & 10-Q financial statements for 2,600+ NASDAQ companies.
Discussion
Questions or observations about this dataset? Join the conversation on X.
No comments yet. Start the conversation on X.