Bunpro Full Sentence Trainer: practice full-sentence output with feedback

[jirachi]

Hey there,

I’ve struggled a lot with starting with output in the past, partly because just coming up with sentence ideas on my own for practice is actually a pain after a while. I don’t want to fill in blanks forever, but there also doesn’t seem to be anything out there that forces me to translate whole sentences with proper grading.

This tool pulls all example sentences from grammar points that you’ve studied on Bunpro, presents one in English for you to translate to Japanese in full, grades your translation out of 10, gives you a “correct”, “close” or “incorrect” verdict complete with feedback, and a prompt to type the correct answer after said feedback.

If your translation was not correct/natural enough, then the tool will show it to you again 7 minutes (can be changed) later. Grading is done by GPT-5.4 mini via the API (not standard chatGPT). $5 USD will get you around 2000 sentences graded. It’s meant to work with the free API but for some reason calls weren’t getting responses without spending the minimum $5 during my testing. Grading is done by Gemini 3.5 Flash. This is a free model with usage limits. Other models can be used.

The Github repo has full beginner friendly instructions to get it up and running if you’d like to try. Those with a bit of tech experience will be up and running in little time.

This has been a really useful tool to start building my confidence in assembling sentences from scratch. Hopefully it can be helpful to you too.

If you despise AI with every inch of your being, that’s OK. Just don’t use this. I hate things about it too, but this is an extremely low energy consumption, very productive use case for us Japanese learners.
Perhaps Bunpro could give us something like this on the site natively one day.

Github repo for download

[davconde]

This is a really cool idea! Thank you for contributing it

For some people (myself included), a big part of the AI hate is how dependent it is on big techs and their scummy practices. I think AI has a lot of value on tasks like language learning but I hate giving money to OpenAI and alike.

For things like this we don’t really need huge models, and in my case I get consistently good results for language learning by using Gemma 4, completely self hosted on my laptop (running on NPU with Lemonade Server).

Taking a look into your code, I see that you’re using the new OpenAI Responses API, which is their attempt of having us vendor locked into their ecosystem. Would it be possible for this project to switch to the old Chat Completions API that works with any provider? I’d be happy to send a contribution, or testing if it helps

[jirachi]

Thanks for the response!
Totally understandable. I’ve just pushed a branch that should be Chat Completions compatible.
llm-provider-support branch
Check the .env.example and see if it you’re able to get it to work with your setup. Contributions are welcome if not!

I’m an embedded software guy, so I rarely touch JS/Node. As I’m sure you noticed, this whole thing is definitely not written by me! That said however, I’m happy to properly maintain and test it.

I’ll merge into main if Lemonade Server is working with it. gpt-5.4 is backwards compatible with Chat Completions by the looks of it so for now I’ll leave that as the default. If there are any other free alternatives that you think would be a good fit here (native level Japanese) then please let me know! It’s good to hear that Gemma 4 is delivering you good results. I’m not good enough at the language yet to know if whatever the open models are telling me is actually correct so I’ve been avoiding them for a while.

[davconde]

Thanks for addressing it! And don’t worry, I’m also a C/C++ plus Python guy, so I cannot judge any web code haha

I’ll test it later and let you know, or open a GitHub issue if there’s anything wrong to not clutter the forum.

[jirachi]

Great to hear.

I’ve actually merged the those branch changes into main already and made the default provider Gemini Flash so that nobody here has to worry about paying anything. Both Flash and GPT work fine with the Chat Completions API! Fingers crossed your local setup should have no problems either.

I saw from your Github profile that you’re in Tokyo - hope you have a lovely day! 素敵な一日を！