Tool RPGM SLR Translator - Offline JP to EN Translation for RPG Maker VX, VX Ace, MV, and MZ

[RenderedFatality]

Theres some fancier things you can do with that particular model to speed it up, ik_llama.cpp and a specific argument

--override-tensor "([0-2]).ffn_.*_exps.=CUDA0" --override-tensor "([3-9]|[1-9][0-9]+).ffn_.*_exps.=CPU" or something like that, + tuning --ngl argument. + --fmoe on ik_llama.cpp

I was taking a better portion of an hour earlier capturing responses and creating a tiny dataset to finetune the 4b and 14b earlier, and the 4b is fairly promising, accidently baked it for 8 epochs on 405 examples so is it overfitted? absolutely. but it definitely wasnt bad.

 

[RenderedFatality]

Put together a few fine tunes on some of the Qwen3 models that should be applicable to people from a variety of hardware levels. They should all be useful for folks on constrained hardware. They only were utilized in batchsize 500 or less.

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

You will need to manually assign a jinja chat template to all of these when using LM Studio or VLLM. I never put them into the tokenizer_config.json.

You must be registered to see the links

 

[Shisaye]

Theres some fancier things you can do with that particular model to speed it up, ik_llama.cpp and a specific argument

--override-tensor "([0-2]).ffn_.*_exps.=CUDA0" --override-tensor "([3-9]|[1-9][0-9]+).ffn_.*_exps.=CPU" or something like that, + tuning --ngl argument. + --fmoe on ik_llama.cpp

I was taking a better portion of an hour earlier capturing responses and creating a tiny dataset to finetune the 4b and 14b earlier, and the 4b is fairly promising, accidently baked it for 8 epochs on 405 examples so is it overfitted? absolutely. but it definitely wasnt bad.

Put together a few fine tunes on some of the Qwen3 models that should be applicable to people from a variety of hardware levels. They should all be useful for folks on constrained hardware. They only were utilized in batchsize 500 or less.

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

You will need to manually assign a jinja chat template to all of these when using LM Studio or VLLM. I never put them into the tokenizer_config.json.

You must be registered to see the links

Sounds good, and feel free to use this thread and the anime sharing thread to share models give advice and stuff, but I think I'm done for now.
I was just trying to find an example that shows that DSLR works if you use it with a good model, so that I no longer get bug reports claiming that it doesn't and me not really being able to prove to them that it does, because they won't pay DeepSeek for it.

My focus is now back to improving parsing, plugin compatibility, etc., since I'm not actually planning to use a local LLM.
My hardware is so terrible I definitely spend more money on power and AC using them, than I would paying for DeepSeek during their off hour discount.
The only reason I was able to test so many models is because most already failed at the second batch.
The actual full test with the qwen3 model now took me 9.5h and that was only a single game.

 

[Shisaye]

I've released v1.141
Fixed an issue with the DSLR wrapper that sometimes destroyed \N[] commands at the start of cells.
Fixed also some other stuff concerning \n[] commands that made translation quality worse.
There was no issue with SLR, so if you don't use DSLR you can ignore this update.

 

[Shisaye]

Something I didn't think about much when designing SLR Translator, because it wasn't really necessary when everyone was just using SLR is user friendly debugging.
For example that \N[] error just now was completely invisible to the user because the error already happened in the wrapping stage which means it wasn't actually fully sent to the DSLR engine and as a result the error detection also did not reject those translations since they technically weren't wrong, and the average user looking at the result would have thought it's just the LLM screwing up.
Because since no error was detected by DSLR it also did not log the original and translation texts in the F12 console.

The way I found out was directly placing alerts in the code for those tests, that then showed me that the LLM wasn't actually sent half of the \N[].


To get to the point of writing this, should I make more "built in" debugging stuff? In the sense that it either logs even more stuff in the console by default, even if no error was detected? (That could get really messy and potentially eat a lot of ram)
Should I in general log more stuff directly inside the batch-translation, pre-processing, and post-processing windows?

I also don't want to give information overload that might make it seem like something went wrong when it didn't, or make someone miss an actual error because it's buried inside hundreds of logs.

 

[wibuntr]

i'm always using the first SLR but what is other SLR and DSLR?

 

[Shisaye]

View attachment 5018056
i'm always using the first SLR but what is other SLR and DSLR?

The heck?
There should only be one SLR entry.
Red Google and Red Piggyback shouldn't be in that list either.
Maybe you need to delete the profile folder in Appdata, because there's some wrong old information in there.

DSLR is my universal LLM engine. It's designed to be used with DeepSeekv3, but can also be used with free locally hosted LLMs like qwen3-30b-a3b-abliterated-erotic-i1-IQ3_XSS as well.
It's mostly for people trying to make an AI assisted proper translation, it's way too slow for a quick MTL.