datalyes/DAPFAM_patent

DAPFAM dataset

What’s new (Sept 2025) — DAPFAM patent family retrieval tasks are now in MTEB. 18 tasks (ALL / IN / OUT × 3 query views × 3 target views) are available, including the 6 main ones used in our paper. You can benchmark any model with a single script and reproduce the paper’s results by selecting the same encoder (Snowflake/snowflake-arctic-embed-m-v2.0). Our paper used int8 quantization for hardware reasons; results may vary very slightly (not significantly) if you run in float16/32.

DAPFAM — A Domain‑Aware Family‑level Dataset to benchmark cross‑domain patent retrieval

License: CC‑BY‑NC‑SA‑4.0
Tasks: text‑retrieval (patent family prior‑art retrieval)
Languages: English (eng‑Latn)
Evaluation date span: 1964‑06‑26 → 2023‑06‑20
Cite: Ayaou et al., 2025 — DAPFAM: A Domain‑Aware Family‑level Dataset to benchmark cross‑domain patent retrieval (arXiv:2506.22141)

---

Summary

DAPFAM provides 1,247 query patent families and 45,336 target families with citation‑based relevance and explicit domain labels (IN/OUT). Each positive pair is IN‑domain if query and target share at least one IPC3 code, OUT‑domain otherwise. Text is at family‑level full text (title, abstract, claims, description). Supports both document- and passage‑level retrieval.

What makes DAPFAM different?

• Explicit domain partitions (IN vs OUT) → enables true cross‑domain evaluation.
• Family‑level aggregation → reduces cross‑jurisdiction redundancy.
• Compute‑aware → Small enough to support passage level experimentations on consumer-grade hardware.

---

Benchmark DAPFAM on MTEB

18 retrieval tasks have been added (ALL / IN / OUT × 3 query × 3 target field views). Six of them were directly evaluated in the paper.

Task naming scheme

• Query view: TA (Title+Abstract) or TAC (Title+Abstract+Claims)
• Target view: TA, TAC, or FullText (adds description)
• Subsets: ALL, IN, OUT (IPC overlap filtering)

Task list (18 total)

ALL

• DAPFAMAllTitlAbsToTitlAbsRetrieval
• DAPFAMAllTitlAbsToTitlAbsClmRetrieval (in-paper)
• DAPFAMAllTitlAbsToFullTextRetrieval
• DAPFAMAllTitlAbsClmToTitlAbsRetrieval
• DAPFAMAllTitlAbsClmToTitlAbsClmRetrieval (in-paper)
• DAPFAMAllTitlAbsClmToFullTextRetrieval

IN

• DAPFAMInTitlAbsToTitlAbsRetrieval
• DAPFAMInTitlAbsToTitlAbsClmRetrieval (in-paper)
• DAPFAMInTitlAbsToFullTextRetrieval
• DAPFAMInTitlAbsClmToTitlAbsRetrieval
• DAPFAMInTitlAbsClmToTitlAbsClmRetrieval (in-paper)
• DAPFAMInTitlAbsClmToFullTextRetrieval

OUT

• DAPFAMOutTitlAbsToTitlAbsRetrieval
• DAPFAMOutTitlAbsToTitlAbsClmRetrieval (in-paper)
• DAPFAMOutTitlAbsToFullTextRetrieval
• DAPFAMOutTitlAbsClmToTitlAbsRetrieval
• DAPFAMOutTitlAbsClmToTitlAbsClmRetrieval (in-paper)
• DAPFAMOutTitlAbsClmToFullTextRetrieval

Quick start — run all tasks

import mteb
from sentence_transformers import SentenceTransformer

model_name = "Snowflake/snowflake-arctic-embed-m-v2.0"
model = SentenceTransformer(model_name, trust_remote_code=True,
                            model_kwargs={"torch_dtype":"float16"}).cuda().eval()

task_names = [
  # ALL
  "DAPFAMAllTitlAbsToTitlAbsRetrieval",
  "DAPFAMAllTitlAbsToTitlAbsClmRetrieval",
  "DAPFAMAllTitlAbsToFullTextRetrieval",
  "DAPFAMAllTitlAbsClmToTitlAbsRetrieval",
  "DAPFAMAllTitlAbsClmToTitlAbsClmRetrieval",
  "DAPFAMAllTitlAbsClmToFullTextRetrieval",
  # IN
  "DAPFAMInTitlAbsToTitlAbsRetrieval",
  "DAPFAMInTitlAbsToTitlAbsClmRetrieval",
  "DAPFAMInTitlAbsToFullTextRetrieval",
  "DAPFAMInTitlAbsClmToTitlAbsRetrieval",
  "DAPFAMInTitlAbsClmToTitlAbsClmRetrieval",
  "DAPFAMInTitlAbsClmToFullTextRetrieval",
  # OUT
  "DAPFAMOutTitlAbsToTitlAbsRetrieval",
  "DAPFAMOutTitlAbsToTitlAbsClmRetrieval",
  "DAPFAMOutTitlAbsToFullTextRetrieval",
  "DAPFAMOutTitlAbsClmToTitlAbsRetrieval",
  "DAPFAMOutTitlAbsClmToTitlAbsClmRetrieval",
  "DAPFAMOutTitlAbsClmToFullTextRetrieval",
]

tasks = mteb.get_tasks(tasks=task_names)
results = mteb.MTEB(tasks=tasks).run(
    model,
    output_folder=f"mteb_res/{model_name}",
    encode_kwargs={"batch_size": 16, "prompt_name": None}
)
print(results)

To reproduce the paper’s reported MTEB-compatible results, restrict to the six in-paper tasks listed above. The encoder was run with int8 quantization in the paper; float16 runs on GPU may differ slightly.

---

How to Load the Dataset

from datasets import load_dataset

dc = load_dataset("datalyes/DAPFAM_patent", "corpus")      # 45,336 targets
dq = load_dataset("datalyes/DAPFAM_patent", "queries")     # 1,247 queries
dr = load_dataset("datalyes/DAPFAM_patent", "relations")   # qrels: all/in/out

Counts

• Queries: 1,247
• Targets: 45,336
• Qrels (all): ≈49,869 (positives + sampled negatives)
• Positive qrels: IN ~19,736, OUT ~5,193

---

Evaluation choices

• Metrics: NDCG@100 (primary), Recall@100 (secondary).
• Document-level views in MTEB; paper also explores passage-level retrieval and RRF fusion separately.
• Encoder: Snowflake/snowflake-arctic-embed-m-v2.0; in-paper runs quantized to int8 for efficiency.

---

Citation

@misc{ayaou2025dapfamdomainawarefamilyleveldataset,
  title={DAPFAM: A Domain-Aware Family-level Dataset to benchmark cross domain patent retrieval},
  author={Iliass Ayaou and Denis Cavallucci and Hicham Chibane},
  year={2025},
  eprint={2506.22141},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2506.22141},
}