---
configs:
- config_name: pool_a
  data_files:
  - split: pool_a
    path: pool_a/pool_a.parquet
- config_name: academic_public
  data_files:
  - split: test
    path: academic_public/test.parquet
  - split: train
    path: academic_public/train.parquet
- config_name: assignee_patent_match
  data_files:
  - split: corpus
    path: assignee_patent_match/corpus.parquet
  - split: qrels
    path: assignee_patent_match/qrels.parquet
  - split: queries
    path: assignee_patent_match/queries.parquet
- config_name: assignee_type
  data_files:
  - split: test
    path: assignee_type/test.parquet
  - split: train
    path: assignee_type/train.parquet
- config_name: cpc_section
  data_files:
  - split: test
    path: cpc_section/test.parquet
  - split: train
    path: cpc_section/train.parquet
- config_name: cpc_subclass
  data_files:
  - split: test
    path: cpc_subclass/test.parquet
  - split: train
    path: cpc_subclass/train.parquet
- config_name: forward_citation
  data_files:
  - split: test
    path: forward_citation/test.parquet
  - split: train
    path: forward_citation/train.parquet
- config_name: grant_abandon
  data_files:
  - split: test
    path: grant_abandon/test.parquet
  - split: train
    path: grant_abandon/train.parquet
- config_name: grant_year
  data_files:
  - split: test
    path: grant_year/test.parquet
  - split: train
    path: grant_year/train.parquet
- config_name: kogan_value
  data_files:
  - split: test
    path: kogan_value/test.parquet
  - split: train
    path: kogan_value/train.parquet
- config_name: paper_patent_retrieval
  data_files:
  - split: corpus
    path: paper_patent_retrieval/corpus.parquet
  - split: qrels
    path: paper_patent_retrieval/qrels.parquet
  - split: queries
    path: paper_patent_retrieval/queries.parquet
- config_name: paper_patent_retrieval_hard
  data_files:
  - split: corpus
    path: paper_patent_retrieval_hard/corpus.parquet
  - split: qrels
    path: paper_patent_retrieval_hard/qrels.parquet
  - split: queries
    path: paper_patent_retrieval_hard/queries.parquet
- config_name: papers_cocite
  data_files:
  - split: corpus
    path: papers_cocite/corpus.parquet
  - split: qrels
    path: papers_cocite/qrels.parquet
  - split: queries
    path: papers_cocite/queries.parquet
- config_name: papers_cocite_hard
  data_files:
  - split: corpus
    path: papers_cocite_hard/corpus.parquet
  - split: qrels
    path: papers_cocite_hard/qrels.parquet
  - split: queries
    path: papers_cocite_hard/queries.parquet
- config_name: patent_cite
  data_files:
  - split: corpus
    path: patent_cite/corpus.parquet
  - split: qrels
    path: patent_cite/qrels.parquet
  - split: queries
    path: patent_cite/queries.parquet
- config_name: patent_cocite
  data_files:
  - split: corpus
    path: patent_cocite/corpus.parquet
  - split: qrels
    path: patent_cocite/qrels.parquet
  - split: queries
    path: patent_cocite/queries.parquet
- config_name: patent_paper_citation
  data_files:
  - split: corpus
    path: patent_paper_citation/corpus.parquet
  - split: qrels
    path: patent_paper_citation/qrels.parquet
  - split: queries
    path: patent_paper_citation/queries.parquet
- config_name: patent_paper_citation_hard
  data_files:
  - split: corpus
    path: patent_paper_citation_hard/corpus.parquet
  - split: qrels
    path: patent_paper_citation_hard/qrels.parquet
  - split: queries
    path: patent_paper_citation_hard/queries.parquet
- config_name: ppp_paper
  data_files:
  - split: test
    path: ppp_paper/test.parquet
  - split: train
    path: ppp_paper/train.parquet
- config_name: ppp_paper_patent
  data_files:
  - split: corpus
    path: ppp_paper_patent/corpus.parquet
  - split: qrels
    path: ppp_paper_patent/qrels.parquet
  - split: queries
    path: ppp_paper_patent/queries.parquet
- config_name: ppp_paper_patent_hard
  data_files:
  - split: corpus
    path: ppp_paper_patent_hard/corpus.parquet
  - split: qrels
    path: ppp_paper_patent_hard/qrels.parquet
  - split: queries
    path: ppp_paper_patent_hard/queries.parquet
- config_name: ppp_patent
  data_files:
  - split: test
    path: ppp_patent/test.parquet
  - split: train
    path: ppp_patent/train.parquet
- config_name: pqr_paper
  data_files:
  - split: test
    path: pqr_paper/test.parquet
  - split: train
    path: pqr_paper/train.parquet
- config_name: pqr_paper_patent
  data_files:
  - split: corpus
    path: pqr_paper_patent/corpus.parquet
  - split: qrels
    path: pqr_paper_patent/qrels.parquet
  - split: queries
    path: pqr_paper_patent/queries.parquet
- config_name: pqr_patent
  data_files:
  - split: test
    path: pqr_patent/test.parquet
  - split: train
    path: pqr_patent/train.parquet
- config_name: pqr_patent_paper
  data_files:
  - split: corpus
    path: pqr_patent_paper/corpus.parquet
  - split: qrels
    path: pqr_patent_paper/qrels.parquet
  - split: queries
    path: pqr_patent_paper/queries.parquet
- config_name: renewal
  data_files:
  - split: test
    path: renewal/test.parquet
  - split: train
    path: renewal/train.parquet
- config_name: same_initial_assignee
  data_files:
  - split: corpus
    path: same_initial_assignee/corpus.parquet
  - split: qrels
    path: same_initial_assignee/qrels.parquet
  - split: queries
    path: same_initial_assignee/queries.parquet
- config_name: same_initial_assignee_hard
  data_files:
  - split: corpus
    path: same_initial_assignee_hard/corpus.parquet
  - split: qrels
    path: same_initial_assignee_hard/qrels.parquet
  - split: queries
    path: same_initial_assignee_hard/queries.parquet
- config_name: same_inventor
  data_files:
  - split: corpus
    path: same_inventor/corpus.parquet
  - split: qrels
    path: same_inventor/qrels.parquet
  - split: queries
    path: same_inventor/queries.parquet
- config_name: science_uptake
  data_files:
  - split: test
    path: science_uptake/test.parquet
  - split: train
    path: science_uptake/train.parquet
- config_name: us_foreign
  data_files:
  - split: test
    path: us_foreign/test.parquet
  - split: train
    path: us_foreign/train.parquet
- config_name: vc_backed
  data_files:
  - split: test
    path: vc_backed/test.parquet
  - split: train
    path: vc_backed/train.parquet
license: cc-by-4.0
tags:
- patents
- embedding
- benchmark
---

# PaTRepEval — version v0.4

A SciRepEval-equivalent evaluation suite for patent-trained embedding models. Built on USPTO Open Data Portal granted-patent and pre-grant publication tables, plus external sources (Kogan KPSS, Marx & Fuegi Reliance on Science) for the Pool B / Phase 5 tasks.

## Configs

| Config | Family | Task type | Metric | Notes |
|---|---|---|---|---|
| `pool_a` | corpus | — | — | ~1.05M utility patents, year × CPC stratified. |
| `academic_public` | clf | single_label_classification | f1_binary | 2 classes |
| `assignee_patent_match` | prx | retrieval | p@5, p@10, map | queries aggregated by `mean_by_query_id` |
| `assignee_type` | clf | single_label_classification | f1_macro | 3 classes |
| `cpc_section` | clf | single_label_classification | f1_macro | 8 classes |
| `cpc_subclass` | clf | multi_label_classification | f1_macro | 50 classes |
| `forward_citation` | rgn | regression | kendall_tau | 5y window |
| `grant_abandon` | clf | single_label_classification | f1_binary | 2 classes |
| `grant_year` | rgn | regression | kendall_tau | — |
| `kogan_value` | rgn | regression | kendall_tau | — |
| `paper_patent_retrieval` | prx | retrieval | map, ndcg@10 | — |
| `paper_patent_retrieval_hard` | prx | retrieval | map, ndcg@10 | — |
| `papers_cocite` | prx | retrieval | map, ndcg@10 | — |
| `papers_cocite_hard` | prx | retrieval | map, ndcg@10 | — |
| `patent_cite` | prx | retrieval | map, ndcg@10 | — |
| `patent_cocite` | prx | retrieval | map, ndcg@10 | — |
| `patent_paper_citation` | prx | retrieval | map, ndcg@10 | — |
| `patent_paper_citation_hard` | prx | retrieval | map, ndcg@10 | — |
| `ppp_paper` | clf | single_label_classification | f1_binary | 2 classes |
| `ppp_paper_patent` | prx | retrieval | map, ndcg@10 | — |
| `ppp_paper_patent_hard` | prx | retrieval | map, ndcg@10 | — |
| `ppp_patent` | clf | single_label_classification | f1_binary | 2 classes |
| `pqr_paper` | clf | single_label_classification | f1_binary | 2 classes |
| `pqr_paper_patent` | prx | retrieval | map, ndcg@10 | — |
| `pqr_patent` | clf | single_label_classification | f1_binary | 2 classes |
| `pqr_patent_paper` | prx | retrieval | map, ndcg@10 | — |
| `renewal` | clf | single_label_classification | f1_binary | 2 classes |
| `same_initial_assignee` | prx | retrieval | map, ndcg@10 | — |
| `same_initial_assignee_hard` | prx | retrieval | map, ndcg@10 | — |
| `same_inventor` | prx | retrieval | map, ndcg@10 | — |
| `science_uptake` | rgn | regression | kendall_tau | — |
| `us_foreign` | clf | single_label_classification | f1_binary | 2 classes |
| `vc_backed` | clf | single_label_classification | f1_binary | 2 classes |

## Usage

```python
from datasets import load_dataset

# CLF / RGN: train + test parquets
ds = load_dataset('gebhart/patrepeval', name='cpc_section')
ds['train'][0]  # patent_id, title, abstract, label, year_filed, year_granted

# PRX: corpus + queries + qrels parquets
corpus  = load_dataset('gebhart/patrepeval', name='same_inventor', split='corpus')
queries = load_dataset('gebhart/patrepeval', name='same_inventor', split='queries')
qrels   = load_dataset('gebhart/patrepeval', name='same_inventor', split='qrels')

# Pool A base corpus (read once, share across CLF/RGN at eval time)
pool_a = load_dataset('gebhart/patrepeval', name='pool_a', split='pool_a')
```

## Versioning

This release is **v0.4**. Use the matching git tag (`git checkout <version>` after a `git clone`) when reproducing a paper, since later versions add task configs and may refine splits.