Japanese Patent Claim Translator
Paste the claims from a Japanese application and get English claims in US form, with multiple dependent claims rewritten the way US counsel would rewrite them. It's a first draft — review it before it goes anywhere.
English claims
What US counsel filed for this one (US 2019/0217440 A1)
Runs on a Hugging Face Space. Your claims go from your browser straight to the Space and never touch this server — but don't paste anything that isn't published yet. You get a few translations a day.
What the Claim Translator does
A general translator gives you Japanese claims in English. US counsel would restructure them before filing — different dependencies, US claim form, no reference numerals. This model was trained on what counsel actually filed, so it does that as it translates. Every example below is real output.
Multiple dependent claims
A claim that depends on several earlier claims is split into separate claims or tied to a single parent, the way it usually is before a US filing.
Japanese
【請求項3】請求項1又は2に記載の照明装置を備える車両。
English
3. A vehicle comprising the lighting device according to claim 1.
4. A vehicle comprising the lighting device according to claim 2.
US claim form
Preamble, transitional phrase, then the elements — numbered from 1, with no headings or commentary.
Japanese
【請求項1】基板と、前記基板上に設けられた発光素子とを備える照明装置。
English
1. A lighting device comprising: a substrate; and a light emitting element provided on the substrate.
Reference numerals
Drawing reference numerals are dropped, as they are in most US claims. Numbers that actually limit the claim — ranges, amounts — stay.
Japanese
ドライブプーリ（１３）およびドリブンプーリ（１４）に巻き掛けられる金属ベルト（１５）が、…
English
1. A metal belt for a continuously variable transmission, the metal belt being wound around a drive pulley and a driven pulley, …
Claim Translator FAQ
Why not just use Google Translate or ChatGPT?
You can, and you'll get Japanese claims in English: the multiple dependencies, the reference numerals and the Japanese claim structure all come along, and you fix them by hand. This was trained on what US counsel actually filed, so the fixing is part of the translation.
How was it trained?
On 978 real patent families — the same invention filed in Japan first, then in the US. Each example pairs the Japanese claims as filed with the US claims as filed, whole claim set to whole claim set, so the model learned to translate and restructure at the same time.
How accurate is it?
It's a first draft, not a filing. On 209 applications it never saw in training, every output was in US claim form and none had multiple dependent claims. It can still pick the wrong term, drop a number or word a claim differently from counsel — and it won't add the extra claims counsel often files. The full numbers are on the model card.
Is it free?
Yes. It runs on shared GPUs on Hugging Face, which gives every visitor a few translations a day. If you hit the limit, come back tomorrow, or use the Space directly while signed in to Hugging Face for a bigger allowance.
Where do my claims go?
From your browser straight to the Hugging Face Space that runs the model. They never touch this server. Still, don't paste anything that isn't published yet.
Can I run it myself?
Yes. The model is on Hugging Face under Apache 2.0, with code to run it.
Model, training data and full results: qwen3-14b-jp-us-claims-lora on Hugging Face
Next: check the antecedent basis
Translation can leave a "the" in front of an element that was never introduced. Run the English claims through the Antecedent Basis checker before they go out.
Open Antecedent Basis