patent-drafting-bench / README.en.md
wei0616's picture
Sync patent drafting benchmark from GitHub 3c1949f (#1)
e787b8e
|
Raw History Blame Contribute Delete
3.97 kB

patent-drafting-bench

中文 | English

patent-drafting-bench evaluates the patent drafting capabilities of large language models and AI agents through 60 test tasks. Each task requires drafting claims, a specification, and necessary drawings from a technical disclosure, following the specified jurisdiction and output language. The final deliverable is a Word (.docx) document.

The benchmark evaluates whether a model can turn a disclosed technical solution into a complete, well-supported patent application. It covers claim completeness, clarity, scope, and dependent-claim structure; support and sufficiency of disclosure in the specification; and fabrication of technical content. Tasks include general and case-specific rubrics for item-level human review or LLM-as-a-judge evaluation.

Directory structure

patent-drafting-bench/
├── README.md
├── README.en.md
├── data/
│   └── test.jsonl
└── asset/
    ├── PD-0005_input.docx
    └── ... (60 technical disclosure documents)

test.jsonl is UTF-8 encoded and contains one task per line, with 60 lines in total. document_fields.input_file is relative to the dataset root, for example asset/PD-0005_input.docx, rather than to data/. All attachments are included locally; no COS or S3 download is required.

Dataset size and distribution

Technical fields

Tasks are grouped by the IPC section recorded in their ipc field. Each task is counted once.

IPC section Technical field Tasks Share
A Human necessities 5 8.33%
B Performing operations; transporting 11 18.33%
C Chemistry; metallurgy 5 8.33%
D Textiles; paper 5 8.33%
E Fixed constructions 4 6.67%
F Mechanical engineering; lighting; heating; weapons; blasting 5 8.33%
G Physics 13 21.67%
H Electricity 12 20.00%
Total 8 IPC sections 60 100%

Percentages are rounded to two decimal places.

Jurisdictions and output languages

Target jurisdiction Tasks Share
China (CN) 31 51.67%
United States (US) 15 25.00%
Europe (EP) 14 23.33%
Total 60 100%

31 tasks require Chinese (zh) output, and 29 require English (en) output.

Disclosures with and without images

Input disclosure type Tasks Share
With embedded images 15 25.00%
Without embedded images 45 75.00%
Total 60 100%

This count is based on embedded images in the body of each disclosure DOCX. The 15 disclosures with images contain 17 image references in total.

All tasks require drawings in the final document when necessary. Image content is available in the original DOCX files.

Evaluation rubrics

Each task includes general and case-specific rubrics. General criteria assess claim eligibility, unity, clarity, completeness, scope, dependent-claim structure, and support in the specification, as well as specification structure, embodiments, sufficiency of disclosure, wording, and technical consistency. Case-specific criteria assess technical features, parameters, embodiments, and subject matter described in the corresponding disclosure.

Rubric category Total items across all tasks Items per task
General criteria 900 15
Case-specific criteria 77 0–4
Total 977 15–19

Each criterion includes decision conditions, discrete score levels, and a positive integer weight. Score levels are 0, 0.5, and 1; some criteria use only 0 and 1. Evaluators select the applicable level for each criterion and calculate a score out of 100:

Score = 100 × Σ(weight × level score) ÷ Σ(weights)

The scoring instructions specify evidence requirements and deduction attribution. The same defect is not penalized more than once.